Histórico de versões

Notas de versão

O que mudou em cada versão, da mais recente para a mais antiga. A última versão publicada é a v1.14.6.

v1.14.6

Notas técnicas em inglês.

Upgrade note: one additive migration (100: an empty update_state table for the update notice; no existing row is touched). Upgrading from 1.14.5 is now a double-click: the installer finds the existing installation and upgrades it with no wizard pages. The uninstaller now asks what to do with the data — read below before uninstalling.

  • Updates from the interface. When a newer signed release exists, admins see a notice with four choices: Install, Install at 02:00, Remind me in 3 h and Don’t remind me. Both installs ask for a second confirmation. The download is verified against SyncTide’s release signature before anything runs; a backup is taken, the services are stopped, the installer upgrades in place and the services come back. Progress and the result are shown in Configuration → Runtime → Updates, which also lists the current version and lets you cancel a scheduled install.
    • A version newer than the end of your licence’s update period is not offered; if one is installed anyway, the licence is refused with a clear message and no data is lost — upload a renewed licence or reinstall a version from within the period.
    • Hosted (Docker) installations are updated by SyncTide and do not show the notice.
  • The uninstaller asks whether to keep the data.
    • Keep the data (the default, and what a silent uninstall does): the program is removed completely — install folder included — and the database stays in PostgreSQL. A full database dump, the licence, the datalogger files, reports, report templates, photos, backups and settings are kept in %ProgramData%\SyncTide\saved-<date>. Installing SyncTide again finds them and puts everything back, so a reinstall picks up exactly where it left off.
    • Remove everything: the program, its files and the SyncTide database are deleted. If PostgreSQL holds no other database, PostgreSQL is uninstalled too; otherwise it is left for the other software that uses it. This can take up to 10 minutes, and the uninstaller says so.
    • Folders chosen outside the install folder (for example a reports folder on another drive) are never moved or deleted.
    • The nightly backup task and the desktop welcome page are removed as well.
  • Installer. Unattended install and uninstall (/VERYSILENT with the admin user, password and licence on the command line). An old PostgreSQL data folder left by a previous installation is set aside instead of reused with an unknown password, and setup stops with a clear message when it cannot log in to PostgreSQL.
  • The interface opens without a licence so one can be uploaded, and the nightly backup task is really created at install time.
  • Equipment → Communication shows only the connection settings. The site timezone and the holiday calendar were also listed there as plain text fields, duplicating Metadata → Site & forecasting. They are now edited only in Metadata, and saving the endpoint always keeps their latest values.
  • Fixes
    • The Excel summary report failed (error 500) on every installation with data; it works again.
    • The MQTT service no longer restarts in a loop (shown as “Paused”) when there is no licence for it or no broker configured; it waits and starts by itself when one appears.
    • Stopping a service no longer writes a Python traceback to its log.

v1.14.5

Notas técnicas em inglês.

Upgrade note: no database migrations. Drop-in upgrade from 1.14.4. Files set aside earlier only because they had no readings are taken in on the first scan after the upgrade and leave the list of set-aside files.

  • A file with only its header is an empty delivery, not an error. A datalogger that has nothing to report yet (just started, or no readings in the period) sends its column names and no data. Such a file is now recorded as processed with no rows and no longer set aside. A file with lines of data that cannot be read (for example values without a valid timestamp) is set aside exactly as before, and a datalogger that stops sending data still raises its communication alarm.

v1.14.4

Notas técnicas em inglês.

Upgrade note: no database migrations. Drop-in upgrade from 1.14.3. Only Docker installations change; the Windows standalone behaves as in 1.14.3.

  • Docker: the server’s own log is visible with docker logs. Like every other service, the server now writes its log (security-policy reports, report and messaging warnings) to the container output instead of a file inside the container that was lost at each upgrade.

v1.14.3

Notas técnicas em inglês.

Upgrade note: no database migrations. Drop-in upgrade from 1.14.2.

  • Devices kept for their history no longer count as silent. A device whose Communication Lost alarm is turned off (a device kept only for its history, a copy of another installation, a test device) is still listed with its last data, shown as “off”, and is no longer counted among the devices that stopped reporting — in the health panel. The dashboard’s device status already treated it this way.

v1.14.2

Notas técnicas em inglês.

Upgrade note: one database migration runs on first start (099); it only adds a table. Take a backup before upgrading, as always.

  • A new datalogger export is read. Files whose first three lines are a title, the column names and a units line (with the date format, e.g. dd/MM/yyyy HH:mm:ss) are imported using the declared date format, and the units are kept. Every other format is read exactly as before.
  • A file with empty cells no longer breaks the reader. A file with no valid rows is set aside with “No valid rows parsed”, and set-aside files always state the kind of error.
  • The ingestion status is the same on every installation. The last run’s summary is kept in the database, so it is shown even when the ingestion worker runs separately from the server.
  • Read-only accounts see contact details masked in the messaging history and statistics (for example +351•••••789).
  • Ports & firewall: the server now reports the ports it is configured on.

v1.14.1

Notas técnicas em inglês.
  • Windows installer: embedded Python 3.13.16. A full installer upgrade moves the runtime to 3.13.16; the build now refuses to ship a runtime that differs from its pin.
  • FTP and SFTP images pinned to the exact builds already in service.

v1.14.0

Notas técnicas em inglês.

Upgrade note: two database migrations run on first start (097 and 098). They only add tables; they change no existing data. Take a backup before upgrading, as always. After upgrading, the first ingestion cycle reads every file in the intake folders once to rebuild its cache; later cycles only read new or changed files. Every session now ends on the server after the same idle time the screens already use (30 minutes by default): an integration that calls the API less often than that must log in again when it gets a 401.

  • A file that cannot be imported is set aside, not retried for ever. It is listed with its reason in the ingestion status, can be retried with one click, and is retried automatically when it changes or after an upgrade. Temporary problems (database, permissions, disk) never set a file aside.
  • Faster, lighter ingestion and polling. Measurements are written in batches and a slow database no longer holds up the other devices of the same protocol.
  • Faster alarms, charts and forecasts. The alarm engine reads only the new data since its last check, the Explorer filters tags earlier, and forecast training on hourly models reads the hourly summary instead of years of raw samples.
  • Lighter screens. The interface loads in parts (the sign-in page downloads about a fifth of what it did), charts no longer redraw when nothing changed, and long alarm lists scroll smoothly.
  • Security. Sessions expire on the server after inactivity; error messages never show internal details; a Content-Security-Policy is reported; the Windows installer no longer suggests a fixed database password; base images updated (Python 3.13.16) and pinned.
  • Windows services wait for PostgreSQL before starting.

v1.13.1

Notas técnicas em inglês.
  • Forecast cards say which inputs the model really uses. Inputs the last training dropped (for example weather that did not help) are shown apart, as “not used in the last training”, instead of as if they were in use.
  • A device that sends data in batches no longer shows a warning between deliveries. When a short-horizon forecast waits for the next delivery of a device that reports every few hours, the card shows a neutral note with the delivery cycle; a delivery that is really late is still a warning.

v1.12.1

Notas técnicas em inglês.
  • The asset tree greys a silent device again. A device whose last communication is older than its configured communication timeout now shows grey in the asset tree, exactly as on the map and in the alarm engine. Before, a time-zone mismatch kept every device green in the tree.

v1.12.0

Notas técnicas em inglês.

Upgrade note: one database migration runs on first start (095). It only adds a table, three columns, a CHECK constraint and indexes; it changes no existing data. Take a backup before upgrading, as always.

  • A failed message never costs an alarm. If sending a notification hit a database error, the alarm check for that minute could be lost for every device, while the log still reported success. Each notification now runs on its own: an error affects only that message, and it is logged.
  • Alarm rules can wait before they fire. A numeric alarm rule takes an optional delay: it raises only when its condition has held that long (for example, a value at 0 for more than 10 minutes). A single short breach no longer raises it. The alarm delay is available on devices with live data, not on CSV-file devices.

v1.11.0

Notas técnicas em inglês.

Upgrade note: two database migrations run on first start (093 adds tables and columns; 094 converts the records of a licensed add-on, when present, and verifies them before removing anything — it aborts the whole upgrade on any mismatch). Take a backup before upgrading.

  • Safer device settings. Changing a device’s protocol keeps its site time zone and holiday calendar; out-of-range numbers are refused with a clear message instead of a server error.

v1.10.3

Notas técnicas em inglês.
  • Photos attached to records survive a backup and a restore elsewhere. They are stored relative to their folder and the Windows backup now includes them.

v1.10.2

Notas técnicas em inglês.
  • Every edit checks who may touch the device. Editing points, offsetting a period, adding points and reverting an edit now check the user’s device scope, like the forecast fill already did. An administrator restricted to some devices could previously change measurements on any device.
  • Series the platform rebuilds itself are not editable. Some series are recalculated from their own source records, so an edit made in the Explorer was silently overwritten at the next rebuild. They are no longer offered for editing, and reverting one of their records from the edit history is refused; they are corrected at their source.
  • A fill records what was really accepted. The forecast’s generation time and the number of points the operator adjusted are now worked out by the server from the stored forecast, not taken from the browser.
  • Only fill points are joined across a gap. Points added by hand are drawn as before; only points staged or accepted from a forecast are linked into one line.
  • The user manual describes the fill as it works since 1.10.1: the stored forecast, point for point.

v1.10.1

Notas técnicas em inglês.
  • A fill is the forecast, point for point. “Fill gap with forecast” now stages exactly the stored forecast: one point per forecast bucket, at the instant and with the value and band the dashed forecast line shows, so an accepted point sits on that line. 1.10.0 interpolated the forecast onto the device’s sampling grid, which put accepted values beside the line and, on a device sampled every second (OPC UA, Modbus, MQTT), meant thousands of invented points. The device’s sampling now decides only where the gap is, by the same rule the chart uses to break a line, so a forecast point never lands between real samples. The fill bar says “every 1 h” for an hourly model, and staged or accepted fill points are drawn as one line.

v1.10.0

Notas técnicas em inglês.
  • Do not downgrade after the first accepted fill. A 1.9.x build does not know accepted-forecast rows: it would train on them, drop real data that arrives at a filled instant, and remove replaced rows on a revert.
  • Fill a communication gap with the forecast. In the Explorer’s editor a fourth tool, “Fill gap with forecast”, takes a finished gap the operator drags on the chart and stages what the prediction module had stored for it: one point per sampling interval of the device, on the device’s own grid, with the forecast band behind. Any point can be dragged before accepting; accepting needs a reason and makes one batch, “Forecast accepted”, that records the horizon, the model’s training date and how many points were adjusted. It reverts like any edit. Instant tags only for now; totalizers say so. Needs the editor and prediction modules.
  • Accepted points stay out of the model. Training, inference anchors, the band calibration, the arrival-cadence probes, the ceiling measurement and the forecast accuracy panels all leave them out, so the model is never trained on, or scored against, its own output. They are drawn with a dashed ring and the hover says “forecast accepted”.
  • Real data wins when it arrives late. If the logger later delivers the period that was filled, the real samples replace the accepted ones, the history says how many were replaced, and reverting the batch afterwards leaves the real data alone. The ingestion worker now takes the editor’s per-series lock before every write, so an acceptance and a late delivery of the same instants can never cross.
  • Forecasts over gaps are kept longer. Forecasts over instants that were never measured are kept for a configurable number of days (Configuration › Runtime › Retention, “Forecasts over gaps”, default 180, minimum 7) instead of seven, so a gap can still be filled weeks later, and they survive a fill until real data replaces it, so a fill can be reverted and redone with another horizon. The nightly sweep’s probe was rewritten to keep using the measurements index on compressed chunks.

v1.9.8

Notas técnicas em inglês.
  • Raw mode shows a week of a fast logger, and says when it cannot. Raw keeps the newest samples of each series and drops the rest. The Explorer asked for 15 000, which on a device sampling every 8 seconds is a day and a half: a nine-day window drew only its right edge, and the blank left half read as missing data until a zoom made it appear. The Explorer now asks for 100 000 per series (the server accepts up to 200 000), and when a series still had more than that, a line under the chart says how many of its newest samples are drawn and that zooming in or picking an interval shows the whole period. The data table’s footer carries the same mark.
  • Responses are compressed. Every reply above two kilobytes is gzipped. A hundred thousand raw rows were six megabytes of JSON; they are now well under one on the wire, and every other page is lighter for the same reason.

v1.9.7

Notas técnicas em inglês.
  • The confidence band is calibrated against the misses the model has really made. The band is a q10/q90 pair fitted on the training matrix, where every row is a clean “h ahead of its anchor” on a regular grid. Live, a forecast’s miss carries what the matrix never shows the boosters: an anchor hours stale on a twice-daily feed, a catch-up delivery, a day the site behaved unlike any in the history. On the first hosted client’s 7-day flow model the fitted band covered 77%, then 64%, then 59% of the real misses against the 80% it is labelled with — and the label is the only number an operator, or a forecast-alarm rule in cautious or sensitive mode, reads.

    Each inference now pairs the stored forecasts of the last seven days whose instant has passed with what was then measured, takes the 10th and 90th percentiles of those signed misses — the interval that would have covered 80% of the recent past — and widens every new point’s band to contain predicted plus those. Widened only: the fitted band is never narrowed, so a model whose quantiles are already honest is untouched. A one-sided bias widens one side. Below 48 recent misses the band is left as fitted, which is also the state of a new model, and the decision is logged every run.

v1.9.6

Notas técnicas em inglês.

A release made of a general audit: what could stop alarms or data collection without anyone noticing, what could freeze the platform, and four security holes. Migration 090 runs on upgrade (a few seconds).

Alarms and data collection keep going

  • One faulty device can no longer silence every alarm. The alarm engine evaluates all devices in one pass. A database error on a single device used to abort the pass and undo the alarms already recorded in it — for every device — while the engine still reported itself healthy. Each device is now evaluated on its own.
  • Nothing waits forever any more. Database waits, queries and idle transactions have limits; a CSV ingestion run that hangs is stopped and retried; two public endpoints no longer block the whole server while they wait on the database; the nightly retention continues past a failed step.
  • The Windows service waits for the database after a reboot instead of failing with a misleading schema error; the MQTT subscriber no longer restarts (and loses messages) over a transient database error.

Faster as history grows

  • Fewer, larger storage chunks for new data (a week instead of a day). Planning a query costs a step per chunk; with years of history that had become the main cost of the queries the alarm engine runs every 5 seconds.
  • Faster queries for the alarm engine’s latest values, the fleet health and flow views, the fleet summary and the prediction ceiling card, and new indexes for the alarm list. Results are identical — every rewritten query was compared with the old one.

Security

  • A Telegram reply acknowledges an alarm only from the chat the alert was sent to.
  • VPN profiles: command directives written as --up or inside a <connection> block are refused like any other.
  • The tag limit can no longer be bypassed by two simultaneous requests for the last tag.
  • The Telegram bot token no longer appears in delivery errors; the SMS webhook secret is shown to administrators only; operators restricted to some sites no longer read other sites’ mappings, monitoring settings or PLC addresses; alarm acknowledgement by token is POST-only and rate-limited; a licence is verified before it replaces the current one; baseline security headers on every response; passwords set by an administrator need at least 8 characters.

Licences

  • MAC-bound licences work in virtual machines. VMware and Hyper-V network adapters were ignored, so a licence could never match inside a VM.

Web interface

  • When the server does not answer, the interface says so. The top bar shows OFFLINE since hh:mm and the alarm list keeps its last known state — it used to empty itself, which read as “no alarms”. Polling backs off while the server is down and pauses in background tabs.
  • An expired session returns to the login page with a message; a failed start shows an error with a retry; a failed chart refresh keeps the last data.
  • Moving the mouse over a chart with a lot of data no longer freezes the page.
  • Fonts and the map library are served by SyncTide itself — nothing is loaded from Google or unpkg. Only map tiles come from OpenStreetMap.

Installer

  • Licence agreement 2.0 in English and Portuguese; the installer is available in Portuguese.
  • THIRD_PARTY_NOTICES.txt lists every bundled component with its licence.

v1.9.5

Notas técnicas em inglês.

The tag limit counts what you use

  • Only active mappings count toward the licence. The limit counted every column a datalogger sends, mapped or not. A CSV logger that reports eighteen columns, of which six feed the platform, needed a plan for eighteen. It now counts equipment tags mapped to a system tag with the mapping switched on — the same six.
  • Only those tags keep history. What is stored is exactly what is counted: a column without an active mapping still appears in the tag list (so it can be mapped) and still shows its current value in the Real Time tab and in the device’s freshness, but its values are not written to history. Mapping it later starts its history from that moment. For CSV devices the files still on the server can be re-read to fill it in; live protocols have no copy to re-read. The Active switch in the mapping editor says so.

Lines break only where data is missing

  • A coarse interval no longer reads as a communications failure. A line broke whenever two points were further apart than one and a half times the device’s sample period. At a 30-minute interval over 15-minute samples every pair of points is 30 minutes apart, so every pair was a “hole” and the Explorer drew a scatter of dots. The threshold is now the wider of the sample period and the chosen interval, in the Explorer and in both dashboard trends — the rule the Prediction page already used. An interval with no data still breaks the line.

Licences on hosted instances

  • A licence replaced by the platform operator takes effect on its own, without restarting the services — every process notices the file changed.
  • Hosted instances no longer offer an upload that cannot work. The licence there is installed by the operator; the License tab says so instead of showing an upload box that failed with a file-system error. Failed actions now show a red notice instead of a green check.

v1.9.4

Notas técnicas em inglês.

Years of history now count

  • The long-horizon class trains on up to three years (it was capped at 365 days). One year shows each season exactly once, so the model learned what this September looked like rather than what Septembers look like; two years is the minimum for a season to be an average instead of a memory, and three keeps an odd year from deciding it alone. The cap is a ceiling, not a requirement: a client with 77 days trains on 77 days, and a client who keeps years of history is the reason to keep them. Weather archive backfill and retention widen with it automatically. A validation probe on a signal with annual structure asserts that the extra years never cost accuracy.

The ceiling card reads the season too

  • Two seasonal references on “Measure the ceiling”: repeat the value from 52 weeks ago (the same hour and weekday one year back — 364 days so the weekday lines up) and the average for this week of the year and hour, a table of the season across the years on record. Both show a dash until the history can honestly answer them — a year for the first, two years for the table, because with one year every cell would be a single remembered day and the number a memory, not a forecast. The card now scores up to 400 days so the first has a year to pair against.
  • The reading sentence names phases, not a distance. It said “repeating the last value beats the table up to about 24 h ahead”, as if there were a line. On a signal with a daily cycle there is not: an hour ago and a day ago both beat the table while six hours ago — the opposite phase of the day — loses badly. It now lists which lags win and which lose, and says why.

v1.9.3

Notas técnicas em inglês.
  • The delivery-cadence guard reads the clock it should. 1.8.5 taught the staleness guard to recognise a feed that uploads a few times a day by measuring the gap between its deliveries. It measured that gap inside a window of sample times, which is the wrong clock: a delivery is a moment rows arrive, whatever hours they carry. On the first hosted client the logger went quiet from 26 to 29 September, then delivered a three-day-old catch-up file at 15:17 and a full day of fresh data at 21:56 — two arrivals, hours apart. Inside a 26-hour window of sample times only the fresh one existed, the guard saw a single arrival, found no cadence, and fell back to two buckets: “latest value is 681 min old (limit 120 min)”, the exact badge 1.8.5 was written to retire, back on the model for ten hours of every day. The window is now seven days of arrival time (bounded on sample time for the index), so a catch-up delivery counts as the delivery it is, and a feed silent for longer than a week has no recent cadence and is judged dead by the bucket rule — the right verdict for one that has really stopped.

v1.9.2

Notas técnicas em inglês.

Measure the ceiling before you train

A new button on every tag’s prediction panel, Measure the ceiling, scores the trivial references on the tag’s own history — nothing is trained:

  • predict the long-run average
  • the average for this hour of day, and for this weekday and hour
  • repeat the value from 1 h, 6 h, 24 h and 7 days ago

Each is an error in the tag’s unit. Read together they answer, in ten seconds, what a model card cannot: whether the signal moves at all, up to what distance recent history beats the calendar, and the number a horizon has to get under to be worth having. The card says it in words: how much of the miss the hour of day alone removes, and where the last value stops beating the table.

Why. On the first hosted client the shipped 7-day flow model scored 2.46 where the hour-of-day table scored 1.64 — beaten by 33% by a table nobody trained — under a green “beats baseline” badge, because the baseline was persistence at seven days, which anything beats. Two more tags carried models on signals that never moved (a battery voltage, a meter’s housing pressure, R² of −4333). This table would have shown all three before anyone pressed Add horizon.

A signal whose spread is under 1% of its level is called out as flat: there is nothing to forecast, and a plain threshold alarm is the honest tool.

Any operator may measure; the card is read-only. The two averages are scored on the history they are built from and say so — slightly optimistic, which is still the right bar. Route: GET /predictions/ceiling/{device_id}/{tag_id}, gated by the prediction module.

v1.9.1

Notas técnicas em inglês.

The accuracy panel scores every forecast you were shown

The panel used to score only the model version currently on screen, so a nightly retrain reset its sample to zero at 02:30. On a feed that uploads twice a day the count could reach 19 by midnight — one short of the 20 the “inside the band” figure needs — and then reset. The percentage never showed.

Scoring only the current version was right for the case that prompted it (data replaced, model retrained, the panel reporting the old model’s 31 m³/h miss under a badge saying 4.7) and wrong for the ordinary night, when a retrain on one more day of the same data is the same model. The panel’s own title says what it measures: how this horizon did in the window shown, and the forecasts an earlier version wrote were the forecasts the operator was shown.

Every genuine forecast in the window is now scored. The version split is reported in words instead — how many points came from an earlier version, and what this version has managed on its own — so a retrain that really changed the model is still visible. Hindsight stays excluded.

What seven weeks of one client taught the model

A walk-forward evaluation over the whole history of a hosted client’s flow (train on everything up to each origin, forecast the next seven days, fifteen origins from 6 August to 3 September) settled three questions:

  • Weather never helped at seven days — not at a single origin, even given the realised weather at the forecast instant, which is more than any forecast could know. Average miss 2.29 m³/h with it against 1.93 without. The per-model holdout selection shipped in 1.9.0 is the right mechanism and it drops the columns here on its own; on the demo tenant a one-day counter model keeps them because there they earn their place.
  • The self group hurt at every origin at this horizon (2.25 against 1.93), confirming the 1.9.0 threshold from a second angle.
  • The model does not get better with more data at this horizon. Its miss tracks the difficulty of the week being forecast, not the length of the history: a calendar-only model beats a plain hour-of-day average by about 3% throughout, and both roughly double during the late-August anomaly that nothing could anticipate. Seven days out, the hour-of-day pattern is the signal and the day-to-day residual is not predictable from calendar or weather. That is the ceiling for this kind of tag, and the model sits just above it.

Tried and not shipped: a single slow level feature (3-day or 14-day mean at the anchor) gains about 4% on average but flips sign with the window chosen — a 7-day mean loses, a 7-day median wins — which on two months from one site is the signature of fitting the site, not the problem. Worth revisiting with more sites.

v1.9.0

Notas técnicas em inglês.

Forecasts stop fitting their own noise

A long-horizon model was being beaten by a lookup table. Measured on a live tenant — 57 days of 1-minute flow — the shipped 7-day model scored MAE 2.46 where “the average flow at this hour of day” scored 1.64. A table nobody trained was 33% better than the model.

The cause was the model’s own history. Every model was given the target’s lags out to 24 h plus rolling mean/std/min/max over 6 h and 24 h, all measured at the anchor. That is real evidence while the horizon is short. Once the horizon reaches a full day it is not: the calendar already carries everything a day-old summary could say, and twelve extra columns of it just give the trees somewhere to fit noise. Holdout error with the group and without it, same rows, same hyper-parameters:

horizon instant target counter target
15 min 1.98 vs 2.98 — keep —
60 min 2.51 vs 2.98 — keep 0.40 vs 0.40 — neutral
6 h 1.83 vs 1.88 — neutral 1.93 vs 1.71 — drop
24 h 3.99 vs 1.88 — drop 4.41 vs 1.71 — drop
7 d 2.52 vs 1.65 — drop 6.62 vs 5.88 — drop

At 24 hours the group takes R² negative on both target kinds (−0.62 and −1.52) — the model becomes worse than predicting the average — while dropping it leaves R² around +0.5.

So the group is no longer built at or above a one-day horizon. The threshold is the largest rolling window, not a number tuned to one site: once the horizon reaches the span the group summarises, the summary is a whole window stale relative to what it is being asked about. Short horizons are untouched — below a day the group earns its place, and at 15 minutes it nearly halves the error.

On the live 7-day flow model this moves holdout MAE from 2.52 to 1.65 and R² from 0.35 to 0.65. Every model retrains on its next nightly pass; nothing needs to be reconfigured.

  • A model now records which quantity it was fitted on (migration 089). It used to be read back from the self-feature prefixes, which a long-horizon model no longer has. Without the column the service would fall back to the tag’s current type, and a tag retyped after training would have its forecasts rebuilt as the wrong quantity, silently.
  • No persistence verdict is claimed above the threshold. The column that scores “nothing changes” is a self feature, so it is not there to score — and “nothing changes” is not a forecast anyone makes a week ahead. The badge reads as silence rather than as a bar cleared.

Weather has to earn its place

Enabling Open-Meteo is one switch for the whole installation, so a device gets weather columns because somebody somewhere wanted them — not because this tag’s demand answers to rain. On the same live model, holdout MAE:

columns MAE R²
no weather 1.65 0.65
temperature only 1.81 0.61
temperature + precipitation 2.48 0.41
precipitation only 2.63 0.31

Precipitation is mostly zero with occasional spikes — high leverage, little signal for a water-supply flow — and a young model has too few rows to learn that it means nothing here. That was a third of the accuracy, spent on columns nobody had checked.

The trainer now fits both ways on the same split and keeps whichever scores better on held-out rows, logging the decision either way. Two extra fits per nightly retrain, about a second. It reverses itself the moment the site gives it a reason to: a tag that really does follow the rain keeps its weather on the first night the numbers say so. Nothing to configure, and the weather stays enabled and collected — this only decides whether a given model is fitted on it.

v1.8.7

Notas técnicas em inglês.
  • The accuracy panel scores forecasts, not hindsight. A stored point written at or after the instant it predicts is not a forecast, and its error is not forecast error. The panel already detected those points and named them in red, but still counted them in “average miss” and “inside the band”. On a tenant whose data arrives twice a day that is most of the sample — the model anchors on its newest sample, so the hours between that sample and now are already past when inference runs, and 10 of 12 scored points were hindsight. An operator read “92% inside the band” off it and concluded a correctly calibrated band was broken. Those points are now excluded from the numbers and stated as excluded; when every comparable point is hindsight the panel says so instead of printing a score.
  • A coverage percentage needs a sample. With an honest 80% band, 11 or more of 12 points landing inside happens about 27% of the time. The percentage is now shown only once at least 20 banded points back it; below that the chip reads ”—” and the tooltip gives the count and why.

v1.8.6

Notas técnicas em inglês.
  • Forecast band contains its own forecast. The line and the band come from three independently fitted boosters — an L2 fit for the mean and a q10/q90 pair — and nothing tied them together, so the value could land outside the interval drawn beside it. A live tenant showed “27.12 m3/h, interval 22.30-25.07” on 9 of one model’s 180 stored rows, the worst by 2.05 m3/h. The stored band is now stretched to contain the point estimate, at the same boundary where an inverted pair was already put back in order. The forecast itself is never edited: it is the number the operator acts on and the number a forecast alarm compares against a threshold. The service logs how many rows needed it, next to the existing quantile-crossing count — a model needing it often has a band not worth drawing.

v1.8.5

Notas técnicas em inglês.
  • Equipment metadata: the three save buttons (metadata, location, group) each post only their own fields, but the server wrote every column from the request — “Save metadata” right after “Save location” wiped the coordinates (seen on a hosted tenant the day the prediction module was licensed; the weather predictors silently never arrived), and “Save location” wiped the name and group. The update is now partial: only fields sent change.
  • Prediction chart: measured history is drawn at the model’s bucket, but the line broke at the device’s raw sample period — a 5-minute feed under a 1-hour model came out as loose dots with no line and no band between them. The break now follows the spacing actually drawn.
  • Forecasts for feeds that upload a few times a day: the staleness guard could only recognise batch feeds that deliver more often than every two hours. A client uploading twice a day over FTP was permanently classified as a dead feed — every inference tick skipped, no forecast ever written. The service now also measures the tag’s delivery cadence, and the tolerance it earns is bounded by a quarter of the model’s horizon: twelve-hour-old data is fine for a 7-day forecast and still refused for a 1-hour one.

v1.8.4

Notas técnicas em inglês.
  • Forecast alarms: editing a rule’s threshold or operator while its alarm is open updates the open event at once (the card no longer shows the value the episode opened with), and every later update from the evaluator carries the rule’s current condition too.

v1.8.3

Notas técnicas em inglês.
  • Forecast alarms: when the predicted crossing sits in the first forecast bucket after the last sample, the condition is already true and forecast to hold. The alarms card and the rule list now say “already > 80 m³/h now — predicted to hold for the next 15 min” instead of a “predicted at hh:mm (awaiting data)” that trailed the clock forever. The events feed carries the model’s bucket and the rule’s forward window.

v1.8.2

Notas técnicas em inglês.
  • Forecast alarms: “predicted at 10:53” is judged against the data, not the clock. While the crossing instant is still ahead of the last sample the rule has seen (ingestion lag), the card and the rule list say “awaiting data” instead of “passed” — which contradicted the alarm still being open. The events feed and the rule list carry the evaluation anchor.
  • Predictors: the accumulation kind offered follows the tag’s type — “Counter” for a totalizer, “Rate” for an instantaneous tag. A legacy wrong value stays visible and flagged until changed.
  • Home alarm feed: curve_rules has no severity column since migration 035; 1.8.1 referenced it and the feed returned 500 (caught on the demo, fixed, now covered by a real-schema test).

v1.8.1

Notas técnicas em inglês.

Follow-up to 1.8.0 from the first day on the demo tenant.

  • Active alarms card: a forecast alarm now says when it is predicted to happen (“predicted at 11:01 (in 3 min)”), marks its value as the peak, and flags “forecast stale”. /home/recent-events carries predicted_for, predicted_peak, forecast_stale, and resolves the name and severity of forecast and curve rules (both used to fall back to “warning”).
  • Forecast alarm editor: the selected confidence mode is explained in one sentence under the buttons, direction-aware (“even the LOWEST value in the band is above the threshold” for >); horizons no longer read “++1 d”.
  • Prediction page: hovering a past instant shows the forecast beside the measurement, the miss (measured minus predicted) and whether it fell inside the band; the earlier-version line is hoverable.
  • Predictors: legacy entries without device_id are named again; a digital tag saved as a predictor is named and flagged “ignored by the model” (the service never used it), instead of rendering as “#12”.

v1.8.0

Notas técnicas em inglês.

Forecast alarms (prediction module)

Alarm rules on what the forecast says will happen, not on what a sensor just measured: “tell me if the flow is predicted to exceed 200 m³/h at any point in the next 12 h”. The event carries when (first predicted crossing) and the peak, lives in alarm_events like every other alarm (acknowledgement, history, messaging escalation unchanged), and shows in the Alarms list as kind “Forecast”.

  • Migration 088: forecast_alarm_rules (keyed on the prediction model), six columns on alarm_events, one-open-event-per-rule index.
  • Rule = tag + horizon, operator + threshold, hysteresis, forward window (default the whole horizon; totalizers sum the predicted volume over it), confidence mode (cautious / normal / sensitive), severity, category, acknowledgement; advanced: clear-after-N-runs (3) and cooldown (horizon, at most 6 h).
  • One event per episode, updated in place; clears after N forecast runs without a crossing or as soon as the predicted instant has passed; the rule freezes (event flagged “forecast stale”) when the predictor goes silent.
  • Evaluation is anchored on the tag’s last sample, never on the clock; only a NEW inference run is judged, never the 5 s watchdog tick.
  • UI: Equipment → Monitoring & Alarms → Forecast alarms (list + editor with a live “would fire now” preview and an ingestion-lag warning); Alarms for this tag shortcut on the Prediction horizon card.
  • Messaging: forecast default template (PREVISÃO SyncTide: … previsto > 200 m³/h às 14:00 (daqui a 6 h) — pico 212 m³/h — 80% de confiança); $predicted_for, $lead_time, $predicted_peak, $confidence for custom templates.
  • Routes: GET/POST /admin/forecast-alarm-rules, PUT/DELETE …/{id}, POST …/preview.

Forecast chart

  • Explorer averages the forecast onto the chart’s bucket (no more 1-minute forecast drawn over a 30-minute history); band not drawn at an averaged bucket, and the legend says so.
  • Earlier model versions (7-day retention) are drawn as a lighter line with their own score, instead of being hidden after every nightly retrain.
  • “Compare with actual” → “Show previous forecasts”.
  • A saved predictor the picker cannot offer is named, not shown as “#12”; a digital predictor is flagged “ignored by the model”.

Security (audit 2026-09-07, phase 0)

  • Real client IPs behind Docker: SYNCTIDE_TRUSTED_PROXY_IPS (CIDRs accepted) feeds both uvicorn’s --forwarded-allow-ips and _client_ip; the hosted override defaults it to the Docker address pool. Fixes the tenant-wide login lockout and audit IPs recording the proxy.
  • Dependencies: starlette>=1.3.1 (CVE-2026-48818, -48710, -54282, -54283), urllib3>=2.7.0, idna>=3.15, asteval==1.0.10 (CVE-2026-55244); numpy/scipy/narwhals pinned; setuptools 84 on the build box.

Notes

  • Additive migration; runs on tenants without the prediction module as an inert table.
  • No Windows installer for this version yet: Docker image only.

v1.7.3

Notas técnicas em inglês.

Release date: 2026-08-19 Schema version: 87 (unchanged — no migrations) Previous version: v1.7.2


Backups actually happen now

backup_platform.ps1 has shipped in the install folder for months with no button and no schedule. Using it meant finding a .ps1 under Program Files and opening PowerShell, so in practice nobody had backups.

  • The installer registers a nightly task. “SyncTide Backup”, daily at 02:00, running as SYSTEM so it works with nobody logged in. Re-running the installer overwrites it rather than duplicating. If Task Scheduler is locked down by policy the install still completes and says so in the log.
  • Backup from the interface. A card under Configuration → Retention lists the archives with date and size, downloads them, and takes one on demand — useful before an upgrade.
  • The script refuses to produce an archive it cannot vouch for. It now finds pg_dump where the bundled PostgreSQL actually installs it instead of only on PATH, checks the exit code, and reads the dump back with pg_restore -l before archiving anything. A run that ends without Created archive: did not produce a backup.
  • BACKUP_AND_RESTORE.md ships with the install. It covers taking a backup, scheduling one, verifying it without restoring, and restoring with the timescaledb_pre_restore / post_restore helpers a TimescaleDB restore needs. Until now the installer shipped no documentation at all.

You can see whether the installation is working

A System health card, refreshed every minute:

  • Devices reporting, measured per device against that device’s own “max minutes without data”. Aggregating would let one dead datalogger hide behind the others still reporting.
  • Workers alive, including the case a process supervisor cannot detect: a worker whose process is running but which stopped doing anything.
  • Last backup, flagged when overdue.
  • Disk free — the database and the backups share the volume, and filling it stops both at once.

Problems are named, not summarised: “3 devices without recent data” sends someone looking, “unhealthy” does not.

You get told when there is a new version

The card shows the installed version and, when the machine can reach the internet, whether a newer release exists, with a link. Offline, air-gapped or proxied installs get a quiet “could not check” and nothing blocks.

This does not apply updates. Update packages are still applied by run_updater.bat, which verifies signatures, dumps the database first and rolls back on failure. A one-click self-update whose rollback has never been exercised against a real failure is not something to ship untested.


Upgrade notes

  • No migrations. Schema stays 87.
  • Nothing changes for hosted deployments beyond the health card: backups there are handled by the platform, and the interface says so instead of offering a button that cannot work.

v1.7.2

Notas técnicas em inglês.

Release date: 2026-08-18 Schema version: 87 (unchanged — no migrations) Previous version: v1.7.1


Security

Clears 22 known vulnerabilities across three bundled dependencies. pip-audit against the pinned requirement set is clean after this release.

Package From To Advisories
pillow 12.2.0 12.3.0 13
python-multipart 0.0.24 0.0.31 5
cryptography 46.0.7 50.0.0 4

python-multipart parses multipart form bodies, so it is the one reachable from an HTTP request — it is the reason this release exists rather than waiting.

cryptography moves four major versions because 50.0.0 is the first release fixing all four advisories against it; 48.0.1 and 49.0.0 each fix only part. The jump is large in number and small in surface: SyncTide uses Ed25519 key loading, signing, verification and PEM serialization, stable since cryptography 2.6.

pyOpenSSL is now pinned at 26.4.0. It arrives via the OPC UA library and its own metadata caps cryptography below 47, so leaving it unpinned silently blocks the fix above and a later reinstall would quietly reintroduce the ceiling.

Build tooling: three high-severity advisories in the frontend build chain (nanoid, postcss, vite) were resolved with lockfile-only changes. The produced bundle is byte-identical.


Upgrade notes

  • No migrations. Schema stays 87; upgrading from v1.7.0 or v1.7.1 touches no data and changes no behaviour.
  • No application code changed in this release — only dependency versions.

v1.7.1

Notas técnicas em inglês.

Release date: 2026-08-18 Schema version: 87 (unchanged from v1.7.0 — no migrations) Previous version: v1.7.0


Why this release exists

The v1.7.0 installer shipped without its bundled prerequisites. It was 123 MB against the ~1030 MB of every release since v1.4.0: PostgreSQL, LibreOffice, pgAdmin, Python, the VC++ redistributable and the TimescaleDB payload were all absent. On a machine that already has PostgreSQL it would have installed; on a clean one there was nothing to fall back on.

v1.7.0 has been marked a pre-release and v1.6.2 restored as the current download while this build replaces it. If you installed v1.7.0 on a clean machine, reinstall with this version. An existing working install is unaffected — the missing pieces are prerequisites, not application code.


Fixes

The timestamp-timezone field said UTC when nothing was set

The equipment Communication tab rendered “UTC” whenever a device’s timestamp timezone had never been chosen. An unset field is not UTC: incoming CSV timestamps are read in the server’s own timezone. On a machine set to Portuguese time that is an hour off for half the year, and the form asserted the opposite.

Worse, opening that dropdown and picking “UTC” — the value already displayed — wrote the setting and shifted every subsequent import by an hour. An action that looked like a no-op was not.

An unset field now says so, and nothing else changes: existing installs keep reading timestamps exactly as before.

The timezone picker

  • Reads like a normal one: “(UTC+01:00) Madrid/Paris/Berlin” instead of a bare IANA name, sorted by offset. Zones that behave identically all year are listed together; the offset shown is standard time, so a label does not appear to change by itself twice a year.
  • Fixed offsets, for a logger whose clock was set by hand and never adjusts for daylight saving. A named city is the wrong answer for those: the same reading is read an hour apart in summer and winter, putting a one-hour seam in the series at every changeover — measured as a 65-minute hole for a logger sampling every 5 minutes.
  • A zone stored but no longer listed still displays as itself rather than silently showing a different one.

The build can no longer produce an installer like v1.7.0’s

synctide.iss marks every bundled prerequisite skipifsourcedoesntexist, so a missing folder yielded a smaller installer and a successful build. The build now refuses to continue when a payload is missing, and refuses to publish an installer implausibly small for the payloads that went into it. The same failure cost us the TimescaleDB bundle across v1.15.0–v1.15.3.


Upgrade notes

  • No migrations. Schema stays 87; upgrading from v1.7.0 touches no data.
  • Nothing changes for an install without the prediction module.

v1.7.0

Notas técnicas em inglês.

Release date: 2026-08-14 Schema version: 87 (migrations 081 → 087 required) Previous version: v1.6.2


What’s new

Prediction module (licensed add-on: prediction)

Forecast any numeric tag from its own history, other tags and the weather.

  • Per-tag, opt-in. Pick a tag, pick one or more horizons from 5 minutes to 7 days. Each (device, tag, horizon) gets its own model — a 15-minute forecast and a 24-hour forecast are independent models, not one stretched.
  • What the model reads: the tag’s own recent values and rolling statistics, other tags you choose as predictors (including tags on neighbouring equipment), the calendar of the instant being predicted (hour of day, day of week, weekend, month, public holidays), and — optionally — the weather forecast for the window ahead.
  • Weather comes from Open-Meteo (free, no API key, per-device coordinates). Measured on a rain-driven runoff signal, adding it improved R² by +0.39.
  • Holidays are computed offline for 250 countries including Portugal’s district holidays, plus a ponte rule for a working day walled in by non-working days. No internet needed — a calendar is a base feature, not an optional one.
  • Totalizer tags forecast VOLUME, not the counter reading. Predicting a counter’s level is the one thing gradient-boosted trees cannot do (measured: R² −1.90 against +0.22 for “assume nothing changes”). The model predicts the volume in each bucket instead, so the chart answers “how much will pass in each quarter hour” and the bars can drive alarms.
  • Confidence band from quantile models, shown as a shaded interval.
  • Honest accuracy reporting. Metrics are measured on a chronological held-out tail, never a random split. Every model is also scored against the naive “nothing changes” forecast, and the UI says plainly when the model loses. When a tag barely varies — where R² is mathematically uninformative — the interface leads with relative error instead.
  • Forecast vs actual. Compare what was predicted against what happened, with the error in the tag’s own unit and how often reality landed inside the band. Points written by an earlier version of a model are excluded and counted, so a retrained model is never judged by its predecessor’s output.

New SyncTidePredictor service (Windows service and Docker sidecar). It trains nightly at 02:30 and scores continuously — every minute for short horizons, hourly for long ones. On an install without the module it starts, sees no licence and idles.

Also in this release

  • Module dependencies. Licensed modules can now declare that they require another. The closure is applied when a licence is minted, so an incoherent licence cannot be issued.

  • Deleting source files no longer takes hours. measurements.source_file_id had no index despite being a foreign key; deleting 1000 source files went from 817 s to 75 s (migration 086).

  • Installer no longer leaves a broken install on a dependency failure. The offline dependency step used to wipe the embedded Python, reinstall everything in one resolver pass and only log the exit code — one bad wheel left a working customer with nothing and no message. It now installs wheel by wheel, verifies the result by importing every required module, and if it cannot, stops with a clear message leaving the machine recoverable.

  • In-app updates now carry the module. The delta package includes new wheels and registers new services; previously it shipped a prediction module that could not run.

  • The build fails when the installer fails. It used to print a warning and still report success, so a release with no installer looked complete.

  • SCHEMA_VERSION is now derived from the migrations folder instead of a prompt that only ever added one.

  • install_local.ps1 and repair_local.ps1 accepted only Python 3.14 and rejected the supported 3.13 — repair_local.ps1 silently rebuilt a working virtual environment on every run.


Upgrade notes

  • Seven migrations (081 → 087). All additive and idempotent. Measured at 0.34 s on an 8.3-million-row compressed hypertable.
  • Nothing changes for an install without the prediction module. The endpoints refuse, the interface hides, and the service idles.
  • The installer grows by roughly 50 MB (scipy, bundled for LightGBM).

v1.6.2

Notas técnicas em inglês.

SyncTide v1.6.2 — schema 80 (no new migrations; safe drop-in upgrade from any 1.5.x/1.6.x database).

⚠️ This installer is UNSIGNED — Windows SmartScreen will warn on first run.

Completes the ingestion-reliability work started in v1.6.1.

  • Stability-based upload guard. A CSV is now ingested only after it is identical across two consecutive worker cycles (same size and modification time), instead of merely being “older than 60 seconds”. This closes the last gap in mid-upload protection: a cellular FTP transfer that stalls for over a minute mid-file could previously pass the age check and be ingested as a partial snapshot. Files untouched for 10+ minutes ingest immediately (bulk historical imports don’t wait), and explicit single-file runs (--force/--file) bypass the guard. The one-cycle confirmation adds ~60 s of latency to brand-new files — irrelevant for periodic logger uploads.

v1.6.1

Notas técnicas em inglês.

SyncTide v1.6.1 — schema 80 (no new migrations; safe drop-in upgrade from any 1.5.x/1.6.0 database).

⚠️ This installer is UNSIGNED — Windows SmartScreen will warn on first run.

Reliability release for CSV/file ingestion, born from a real field incident: a data logger whose files were silently dated into the wrong month.

  • Deterministic timestamp parsing. Year-first timestamps with slashes (2026/08/05 09:18:00) and ISO-8601 with T separator (2026-08-05T09:18:00) now have explicit format anchors. Previously these fell through to heuristic auto-detection, which could swap day and month whenever every row in a file was on a day ≤ 12 — data landed months away from where it belonged, while the file was marked as imported. Every timestamp format ever sent by a fleet logger now parses through an explicit, deterministic pattern (verified by re-parsing 538 real customer files against their stored data — zero deviations).
  • Never ingest a file mid-upload. Files modified in the last 60 seconds are deferred one cycle before hashing. FTP uploads over cellular links take long enough that the worker could previously ingest a truncated snapshot of a file still being transferred — including a corrupted final value.
  • Honest skip messages. A file whose identical content previously failed to parse is no longer logged as “already imported” when skipped — the log now says it previously failed and that --force retries it.
  • Loud warning on unknown timestamp layouts. When a file’s timestamp format matches none of the explicit patterns, the ingestion log now carries a prominent per-file warning naming the risk (ambiguous day/month values can misdate rows) instead of a quiet auto-detection note — so a new logger format gets an explicit anchor added before it can bite.

v1.6.0

Notas técnicas em inglês.

SyncTide v1.6.0 — schema 80 (migrations 075–080, applied automatically on upgrade; additive-only, safe on any 1.5.x database).

⚠️ This installer is UNSIGNED — Windows SmartScreen will warn on first run.

  • Explorer polish (all installations). Y-axis labels now use clean 1-2-5 “nice” ticks inside the data range (no more fractional or below-zero labels on counter axes); shared-scale mode no longer pads a non-negative axis below zero.

(Note: an unrelated document previously occupied this filename under the pre-May-2026 version numbering, retired when versions were reset at v1.0.1. See git history.)

v1.5.8

Notas técnicas em inglês.

SyncTide v1.5.8 — schema 74 (no migration; drop-in on 1.5.5–1.5.7).

⚠️ This installer is UNSIGNED — Windows SmartScreen will warn on first run.

Chart bugfix.

  • Totalizer bars no longer overlap. The v1.5.5 time-proportional x-axis broke totalizer bar sizing on unevenly-sampled data: bar width was derived from the average point spacing, so in a dense cluster the bars were drawn wider than their real time gap and overlapped into a solid block. Bar width now tracks the actual (time-axis) spacing — each bar is exactly one bucket wide — so bars stay distinct at any zoom and data distribution. This also makes the “too dense → soft line” fallback trigger off the densest cluster, as intended.

v1.5.7

Notas técnicas em inglês.

SyncTide v1.5.7 — schema 74 (no migration; drop-in on 1.5.5/1.5.6).

⚠️ This installer is UNSIGNED — Windows SmartScreen will warn on first run.

Polish release following the v1.5.5 chart work and the v1.5.6 OPC UA fix.

  • Safer monitoring settings — if the Equipment → Monitoring → Communication settings fail to load, Save is now disabled (with an explanatory tooltip) instead of silently overwriting the device’s real communication timeout and expected data period with defaults.
  • Step-hold and data-gap breaks now compose cleanly — with step-hold ON the Explorer carries the last value across every gap and shows one continuous line (that’s the toggle’s purpose); turn it off to see the expected-period gap breaks again.
  • “Expected data period” only where it applies — the setpoint block appears only on devices positively identified as CSV, instead of also showing for unrecognised protocols.

v1.5.6

Notas técnicas em inglês.

SyncTide v1.5.6 — schema 74 (no migration).

⚠️ This installer is UNSIGNED — Windows SmartScreen will warn on first run.

Bugfix release for OPC UA subscriptions.

  • OPC UA deadband no longer silently kills tags. Some servers — verified on a Siemens S7-1200 exposing data through a Server Interface — reject a monitored item purely because it carries an absolute deadband filter, and report the misleading BadNodeIdUnknown (as if the node were gone), even though a plain read and a filter-less subscription of the same node both succeed. Previously those tags were dropped and collected nothing. The driver now retries any deadband-rejected item without the filter: the deadband stays a best-effort traffic optimisation and is never a reason to lose data. Tags that are genuinely gone are still reported, and a device where nothing subscribes still flips to an error state (no fake “green”).
  • Clearer diagnostics. The rejection log/error no longer asserts a single cause (“Siemens renumbered the ids — use string ids”); it states what the server did and the real possibilities, so a filter-level rejection isn’t mistaken for a missing node.

No configuration changes required. Existing deadband settings keep working where the server accepts them and are transparently ignored where it doesn’t.

v1.5.5

Notas técnicas em inglês.

SyncTide v1.5.5 — schema 74.

⚠️ This installer is UNSIGNED — Windows SmartScreen will warn on first run.

Trend-chart accuracy release: charts now tell the truth about when data arrived — and when it didn’t.

  • Time-proportional x-axis — every trend chart (Explorer, dashboard, equipment) now spaces points by real time instead of sample index. Irregularly sampled data no longer gets visually compressed: an hour is always the same width, and the Explorer axis covers the full window you selected, so a silent datalogger shows as blank space instead of the chart quietly skipping ahead. Hover, drag-to-zoom (including into an empty stretch) and the measurement-editor tools all follow the time axis.
  • Expected data period (CSV devices) — new per-device setpoint in Equipment → Monitoring → Communication: enable “Show data gaps on charts” and enter the cadence your datalogger is expected to deliver (in minutes). When the gap between two samples exceeds the setpoint (×1.5 jitter margin), the trend line breaks — missing data reads as a visible hole, not a straight line pretending nothing happened. Isolated samples between holes stay visible as dots. Purely visual; independent of the communication-loss alarm timeout.
  • Area-fill toggle — new Explorer toolbar button to switch the soft shading under chart lines off/on (per user, remembered across sessions). Bare lines help when many series overlap.

Also includes the fixes landed since v1.5.4:

  • Site map: idle analog tags (e.g. Flow/Level at 0) no longer show as “OFF” in the hover tooltip.
  • Site map: per-device “Map tags” customizer — choose which tags the hover tooltip shows.
  • Explorer: the tag-folder collapse state is now remembered per user.
  • Auth: changing the idle-timeout now applies immediately, without a re-login.

Migration 074 (devices.expected_sample_minutes) applies automatically on first start.

v1.2.3

Notas técnicas em inglês.

Alarm visibility, alarm-history power filters, per-user view persistence and communication-config cleanup. No schema change (schema 66); upgrades from 1.2.1/1.2.2 are drop-in.

Alarms

  • Alarm rows show what actually fired: rule name (e.g. “HH”), the measured value and the comparison (549.7 > 100), plus the device — or the KPI’s name for site alarms. Comms-lost rows no longer show a bogus comparison.
  • The alarm feed updates live (~10 s): new alarms appear and cleared ones drop off without a page reload; acknowledged state follows acknowledged_at.
  • New Cleared filter on Active Alarms. A clear does not waive the acknowledgement: cleared-but-unacknowledged alarms stay in the Active view (with their clear time) until an operator acknowledges them; only cleared + acknowledged events leave it. Badges and “Ack all” count every unacknowledged event; device status pills reflect open alarms only.
  • Alarm history rework: the Type column became Alarm (rule / curve-rule name, with the type beneath); all filters are multi-select (devices, types, status); status filter covers the four lifecycle states (Open/Cleared × Acknowledged/Unacknowledged); custom date range added; the text filter matches alarm names too; CSV export includes the alarm name.
  • /home/recent-events and /alarms/history-all return the rule’s alarm_name + severity for device, site and curve alarm events.

Configuration & Equipment

  • KPI alarm edits can no longer be lost: the tag drawer’s “Save tag” also saves a pending alarm edit, and closing the drawer warns before discarding one. (Previously a threshold change was silently dropped unless the inner “Save alarm” button was used.)
  • Protocol pivot: a device’s protocol can be changed in Communication → Endpoint (e.g. CSV → OPC UA). The form follows the selection, essentials are validated, and the old protocol’s settings are cleaned up on save.
  • Communication forms show only the active protocol’s fields — residue keys from a previous protocol are hidden and stripped on save. CSV shows just folder + timezone (no polling/heartbeat knobs — ingestion is automatic). Hints added to almost every field across all protocols.
  • Escalation Feed lives only in the Messaging Center (removed from the Alarms page) and now shows recipient names, what each message was about, and full date + time.

UX

  • Trees start compact and remember you: the System Tags folder tree and the Explorer (selections, time range, bucket, expanded groups/devices) persist per user and reopen exactly as left; first use starts collapsed.
  • Flow Diagram hides devices without relationships (with a count note) so the layout stays compact.

v1.2.2

Notas técnicas em inglês.

UI polish and data-visualisation fixes across the Dashboard, Explorer and Configuration. No schema change (schema 66); upgrades from 1.2.1 are drop-in.

Dashboard

  • Key Indicators now shows all site KPIs (was capped at 3), with an “Off-target only” toggle to compact the row to KPIs near/outside target.
  • Totalizer KPIs render as bucketed bars; digital KPIs as a step; instant as a line. Site totalizer KPIs (e.g. “Rendimento Totalizadores”) now compute their value over the window anchored on the latest sample, so historical / paused devices still show a value + bars instead of blank.
  • 1Y range added to the Key Indicators (backend window cap raised to 1 year).
  • Sparklines are hoverable — value + timestamp at the cursor.
  • Device cards draw a trend per card tag (line / bars / step by type). Card tags gain a “None” option; setting both to None hides that device’s card.
  • Equipment status pill reflects active-alarm severity and the device’s configured comm-loss timeout.

Explorer

  • Virtual / site-KPI series now align by timestamp with device tags (fixes the virtual series showing blank ”—” against real data); timestamps parsed as UTC; virtual-KPI trend bucketing honours the actual data span.

Configuration

  • Equipment list auto-refreshes (picks up auto-discovered devices) + a manual Refresh button.
  • Virtual-tag expression picker offers only the selected device’s mapped + device-virtual tags; Validate rejects references a device doesn’t have and blocks saving an invalid expression.
  • Home-layout editor shows each repeatable widget’s title; ranking/heatmap widgets are read-only on their variable; heatmap hover tooltip is opaque.
  • Last ingestion card shows the real run timestamp + files found/imported/ skipped.
  • Pie KPI slices are labelled with their values via leader lines.

v1.2.1

Notas técnicas em inglês.

SyncTide v1.2.1 — Web UI fixes & Reports parity

A follow-up to v1.2.0 that addresses operator feedback on the new web interface and restores full Reports parity with the legacy UI. Additive migrations apply automatically on first start (schema 55 → 65); data collection is unaffected.

Fixes & improvements

  • Auto-refresh of new deployments — the SPA shell (index.html) is now served with no-cache headers while hashed assets stay immutable, so a new build is picked up automatically without a manual hard-refresh. New equipment and config changes appear on a normal reload.
  • Trend drag-to-zoom — dragging a time region in the Explorer chart now zooms correctly (the redraw was discarding the in-progress drag state).
  • System tag unit saves — editing a system tag’s unit (e.g. Speed → m/s) now persists; the editor waits for the save to confirm and reports any error instead of closing optimistically.
  • Groups fully configurable — create / edit / delete groups on the Configuration → Groups tab (the stale view was a caching artefact, resolved by the cache-header fix above).
  • Full device deletion — a Delete device action removes a device and all its associated data (cascades through measurements, tag maps, alarms, etc.).
  • Reports — per-tag column processing (legacy parity) — pick the metrics per tag type (Average / Min / Max / Last / Difference) with an ordered selector; the first metric drives the plotted trend (e.g. a totalizer plots Difference). Plus a bucket/interval selector for aggregation.
  • Report History — now shows the Equipment and Interval for each report; the unhelpful Size column was removed.
  • Scheduled jobs — the New scheduled job drawer is now fully wired: real templates and groups, daily/monthly/yearly frequencies, optional email recipients. Editing a job recreates it cleanly.
  • Report Templates — upload new templates, rename / re-type existing ones, replace the underlying file, and delete (soft-delete preserves history). Scheduled reports now email the generated files as attachments via the configured SMTP gateway.
  • System tag editing — editing a non-virtual system tag (e.g. changing a unit) now saves correctly instead of failing validation.
  • Decimals in Explorer — the data table and chart hover tooltip/legend now show at least 2 decimals when a value has a fractional part (whole numbers stay whole), so sub-unit readings are no longer rounded away.
  • Calendar-aligned aggregation — Reports and the Explorer/Dashboard trends gained 1 month, 1 year (and 30 min) buckets, and the hour/day/month/year buckets now align to true period boundaries (full clock hour, 00:00–23:59 day, full calendar month, full calendar year) rather than fixed 30-/365-day windows.
  • Respect the chosen bucket — trends now honour the selected bucket size instead of silently coarsening it for wide time filters; the same period renders identically across “This month” / “This year”. It only coarsens (to the finest feasible size) when a bucket would produce too many points.
  • Per-device timezone — Reports have a Server/Device time basis toggle, and single-device trends, last-communication and alarm timestamps render in the equipment’s configured local timezone.
  • Configurable idle auto-logout — System Configuration → Runtime → Security: sign users out after N minutes of inactivity (0 = disabled).
  • Folder picker — CSV folder paths are now chosen with a real server-side directory browser, never typed by hand.
  • Reports — raw export — choose the CSV columns and their order (Timestamp · Category · Variable · Unit · Value) and the output format (.csv or .txt).
  • Dashboard polish — calendar-style Activity Heatmap, Ranking “Difference (consumption)” aggregate, the trend timeframe selector now belongs to the Asset Tree widget, dead Filters/Export buttons and decorative KPI icons removed.
  • Groups vs Alarm Categories — now distinct. Groups (managed under Configuration → Groups, with parent/child hierarchy, an icon picker and colour mode) are how devices are organised; Alarm Categories classify each alarm rule. Alarm-by-category and messaging escalation route by the firing rule’s category. (Also fixes colour-mode categories rendering as a default icon.)
  • Reworked role permissions — Viewer is now strictly read-only (may change trend filters and pick variables, nothing else); Operator focuses on day-to-day device work and alarm acknowledgement but can no longer edit equipment, configure gateways, change system configuration or manage users; Admin retains full control. Pages and edit controls a role shouldn’t see are now hidden, and every user can change their own name and password.
  • Explorer unit fix — a tag’s live unit (e.g. Speed changed to m/s) now drives the plotted series and axis instead of a stale cached unit.
  • Relationship variable links — when linking two devices (Configuration → Relationships) you can now pin which variable on each side participates: the upstream tag on the source device and the tag it feeds on the target (e.g. a site roll-up virtual tag). The chosen variables show on each link and the source-side tag’s live value renders on the diagram edge.
  • Configurable relationship types with colours — add your own link types (beyond feeds / supplies / monitors) and assign each a colour under Configuration → Relationships → Link types & colours. Renaming a type re-labels its existing links automatically.
  • Flow diagram widget — the Home Flow diagram widget is now a proper colour-coded topology: devices as status-coloured nodes laid out by flow direction, relationships drawn as arrowed edges in their type’s colour, edge labels showing the linked variable / value, and a colour legend. Enable it under Configuration → Home Layout. Parallel links between the same two devices are now fanned apart so their labels no longer overlap, and labels sit on a subtle background for readability.
  • Edit relationships — device links can now be edited in place (pencil icon), not just added/removed.
  • Per-side process on the diagram — each link can choose how its from/to variable is summarised on the diagram: Last value (default), Average, Minimum, Maximum, or Difference over the last 24 h (Difference = consumption for totalizers). The edge shows the from value, then the to value, in the chosen process.
  • KPI alarms — site (cross-device) KPIs can now have alarm thresholds defined directly in the Edit-system-tag drawer (value vs threshold, or out-of-band against the KPI’s Target ± Tolerance), with severity and category. Device-scoped KPIs are still alarmed in that device’s Monitoring & Alarms.
  • Heatmap “Difference” — the Activity Heatmap aggregate now includes Difference (consumption) — per-day last − first — ideal for totalizer tags.
  • Raw export improvements — the exported Variable is always the system tag (device tags not mapped to a system tag are no longer included); values are the raw stored measurements; added Excel (.xlsx) output alongside .csv/.txt; and generated raw exports (CSV/TXT) now preview inline in Reports → History as a table.

Security & robustness hardening (internal audit)

  • Alarm authorization (object-level) — acknowledging/clearing alarms and reading alarm history now enforce per-user device scope. Previously a scoped operator/viewer could acknowledge, clear, or read alarms for devices outside their allow-list by id. Endpoints fixed: /map/alarm-history/{device_id}, /map/alarm-acknowledge/{id}, /map/alarm-acknowledge-device/{id}, /map/alarm-clear-device/{id}, and report-output download scoping.
  • Messaging gateway secrets masked — provider tokens/passwords (Telegram bot token, Twilio auth token, SMTP password, WhatsApp token, router password) are now redacted (***) when gateway config is sent to the browser, consistent with device-config redaction; editing a gateway preserves the stored secret.
  • KPI alarm category persisted — the alarm category chosen for a site-KPI alarm is now saved and reloaded correctly (was silently dropped).
  • Home Layout robustness — the layout editor no longer crashes if a saved layout references a removed widget type, and base widgets can be enabled/disabled (only extra duplicate widgets show a remove button).
  • Report template uploads validated (Excel extension + 25 MB cap).
  • No duplicate alarm notifications — comm-loss and curve-breach alarms now have a database-level “one open alarm” guarantee, so a brief race between the alarm watchdog and the live status view can no longer create duplicate alarms or send operators duplicate messages during an outage.
  • Security audit log — a new append-only audit trail records authentication (login / failed login / logout), account/role changes and password resets, with user, timestamp, IP and outcome. Viewable under Configuration → Audit log (admin). Supports CRA security-logging and NIS2 incident-handling/accountability.

Compliance groundwork (CRA / NIS2)

  • Multi-factor authentication (TOTP) — users can enable two-factor sign-in (Google Authenticator / Authy / 1Password, etc.) under My Account → Two-factor: scan a QR, confirm a code, and store one-time recovery codes. Login then prompts for the 6-digit code; admins can reset a user’s MFA if their authenticator is lost. Directly addresses NIS2 Art 21(2)(j) and CRA access-control requirements.
  • Enforce MFA per role — admins can require two-factor for chosen roles (Configuration → Runtime → Security). Users in a required role are taken through a mandatory MFA-enrolment screen at next sign-in before they can use the platform.
  • Coordinated Vulnerability Disclosure policy (SECURITY.md) — private reporting channel, response SLAs, safe-harbour.
  • Software Bill of Materials — every release now emits a CycloneDX SBOM (SyncTide-SBOM-<version>.cdx.json).
  • Technical documentation templates — EU Declaration of Conformity + cybersecurity risk assessment + CRA Art 14 reporting runbook under docs/compliance/; secure-deployment/hardening guide at docs/HARDENING.md.
  • Optional offline HTTPS — run enable_https.bat to serve SyncTide over TLS using Caddy’s built-in local CA (no public domain, no internet, no paid certificate); HTTP stays available. Reversible with disable_https.bat.
  • Broader security audit log — now also records messaging-gateway create/update/delete, security-policy changes, device connection changes, and all MFA events. Exportable to CSV from the Audit log view.
  • MFA reliability fixes — the login second-factor prompt now appears correctly (a 401 on the login endpoint no longer shows a generic “session expired”); turning MFA off with a wrong current password shows a clear error instead of logging you out; and a role that requires MFA can no longer disable it (the option is hidden) — removing the lock-out risk.
  • Free dependency scanner — tools/security_scan.py (pip-audit + npm audit).

Upgrading

  • Windows: run SyncTideSetup-1.2.1.exe.
  • Additive migrations apply automatically on first start (schema 55 → 65): auth_settings (idle-logout), device_groups + per-rule alarm category, group hierarchy/colour, relationship variable links, configurable relationship types, per-link process, a security audit log, per-user multi-factor authentication, and an optional require-MFA-per-role policy. Data collection from devices is unaffected and continues through the upgrade.

Verifying your download

This installer is not yet code-signed, so Windows SmartScreen may warn on first run (choose More info → Run anyway). Verify integrity:

Get-FileHash .\SyncTideSetup-1.2.1.exe -Algorithm SHA256

Expected installer SHA-256: 1248ac68e972450d074729271976b98c90003f54adde752cc89b7d9694ee240d

v1.2.0

Notas técnicas em inglês.

Released: 2026-05-30 Schema: v55 (no new migrations — drop-in upgrade) Upgrade path: v1.0.x / v1.1.x → v1.2.0 in-place (installer or delta package); cloud docker compose pull && up -d. Bundled stack: PostgreSQL 17.6 + TimescaleDB 2.26.4


Why this release

The operator web interface has been rebuilt from the ground up as a modern single-page application (Vite + React), replacing the previous Streamlit UI. It loads faster, navigates without full-page reloads, and is lighter on the backend. Everything you manage today — dashboard, asset tree, trends, alarms, messaging, reports, equipment configuration — is there, re-laid-out for clarity. Schema is unchanged, so this is a clean drop-in upgrade.


What’s new

New React web interface

  • The whole UI is now a single-page app served by the platform at /ui, fronted by the same Caddy reverse proxy. You still reach SyncTide at the same LAN address (http://<host>) — it lands you on the app and signs you in.
  • Faster navigation (no full-page reloads), cleaner layout, the same feature set.
  • The old Streamlit interface is retired: the SyncTideUI service is no longer installed, and upgrades remove it automatically.

Lighter on the network

  • Live dashboards refresh every device’s latest values in one batched request per tick (new GET /measurements/latest endpoint) instead of one request per device.
  • Screen updates during a refresh tick are coalesced into a single redraw, and the trend cache is bounded — busy multi-device dashboards stay smooth and put far less load on the backend.

European Portuguese

  • Full Português (Portugal) translation, switchable from the top bar.

Under the hood

  • main.py: the FastAPI backend now serves the built SPA at /ui (StaticFiles mount) and adds env-gated CORS for from-source development.
  • New batch endpoint GET /measurements/latest?device_ids=… returns the latest mapped-tag value + timestamp for many devices in a single request (per-device access checks unchanged).
  • Caddy default route now points at FastAPI and redirects / → /ui/ (local, Docker and test configs).
  • Local Windows install is now eight NSSM services (the Streamlit SyncTideUI service was removed).

Operator action when upgrading

  • No database migrations — schema stays at v55. Drop-in upgrade.
  • Windows: run the v1.2.0 installer, or apply the delta package, on top of any v1.0.x / v1.1.x install. The Streamlit service is removed and the new UI is served on first start.
  • Cloud: docker compose pull && docker compose up -d.
  • Data collection from devices is unaffected and continues through the upgrade — only the web interface changed.

v1.0.2

Notas técnicas em inglês.

Release date: 2026-05-11 Schema version: 47 (migration 047 new in this release) Previous: v1.0.1 (same day — robustness work landed in a single sweep ahead of the first customer deploy)


Headline

This release closes the operational gaps left over from the v1.0.1 ship audit. Each item below was either an actual or potential customer-visible problem; none break compatibility.

1. Health monitor — the watchdog of watchdogs

The platform now ships with a self-monitor service that watches the rest of the stack. Every 60 s by default it:

  • Pings the backend’s /version endpoint.
  • Runs docker compose ps and confirms every expected service is Up.
  • After 3 consecutive failures, alerts the operator via Telegram (using the existing messaging gateway — no extra bot needed).
  • On recovery, sends a follow-up “back online” message.
  • Persists state to a new health_monitor_config table (migration 047) so the cooldown survives restarts.

Configurable from Configurations → Health Monitor: enable/disable, Telegram gateway + chat id, check interval, failures-before-alert, cooldown, optional expected-services override.

Disabled by default — turn it on once a Telegram gateway is wired in Messaging Center.

2. Automated daily backups

A new backup Docker service runs backup_cloud.sh daily at 03:00 UTC (BACKUP_HOUR_UTC env var). Captures pg_dump of the SyncTide DB (custom format, TimescaleDB-aware) + tarred named volumes (raw_data / reports / templates / mosquitto / caddy certs) + a sanitised copy of .env. Rolling retention 30 days.

Set BACKUP_RUN_AT_START=1 in .env to take a backup at first boot — useful for verifying the pipeline on a fresh deploy.

The script’s Windows-path bug (named volumes silently skipped under git-bash) is fixed: MSYS_NO_PATHCONV=1 + cygpath -m keeps the docker -v host paths intact.

Restore procedure verified end-to-end against a fresh ephemeral TimescaleDB container: every reference table matches the source row counts (only measurements differs by the rows written between backup and verification — expected, ingest never stops). Full runbook in docs/RECOVERY.md.

3. Security hardening

Five fixes from the v1.0.1 audit:

  • /license/upload now requires admin (was unauthenticated). The Ed25519 signature check inside save_license_file prevents forgery, but the old code let any LAN caller DoS the platform or downgrade-replace with a different valid .lic. Closed.
  • must_change_password enforced server-side. Earlier only the Streamlit UI redirected; direct API callers with a stolen bearer token bypassed the gate. get_current_user now refuses every path except /auth/me, /auth/logout, /users/me/password when the flag is set.
  • init_admin.py seeds new admins with must_change_password=TRUE. A customer who runs docker compose up with the unedited .env.example.docker (SYNCTIDE_ADMIN_PASSWORD=change-me-on-first-login) is force-flipped into the reset flow on first login.
  • _DEFAULT_PASSWORDS expanded to include change-me-on-first-login (the docker template default) so the startup-time scan in main.py catches the case where the seed password was never changed.
  • Bulk alarm endpoints bumped from viewer to operator role. /map/alarm-acknowledge-device, /map/alarm-clear-device, /map/alarm-acknowledge-all, /map/alarm-clear-all. Industrial SOP: viewers see, operators acknowledge. Single-alarm acknowledge stays at viewer level since that’s a routine control-room action.

4. License-manager hardening (audit cleanup)

Zero new code — the audit found the manager already fail-softs on every malformed input we tested (missing, empty, garbage, truncated, bad signature, future-dated, expired, wrong instance id). Backend stays up, all workers idle harmlessly, UI shows a clean banner. Three P3 message-text polish items (clearer “file unreadable” vs “file missing”, less alarmist signature-failure copy) deferred — not customer-facing in normal operation.

5. Operational changes worth knowing

  • Docker image tag bumped to 1.0.2 (compose default updated).
  • Schema version 47 — migration 047 is additive (new health_monitor_config table seeded with a disabled row). Drop-in upgrade from v1.0.1.
  • New documentation: docs/RECOVERY.md — tested restore procedure.

Upgrade

# Pull the new code
git pull
# Rebuild image
docker compose build backend
# Roll the stack
docker compose --profile on-prem up -d

Migration 047 runs automatically on backend startup. The new backup and health-monitor services are created and started in the same up -d.

v1.0.1

Notas técnicas em inglês.

Release date: 2026-05-11 Schema version: 46 (migrations 045 + 046 new in this release) Previous internal: v1.15.4 (versioning reset — see note below)

Versioning note — the pre-launch 1.15.x line was internal-only iteration. v1.0.1 is the first numbered release we ship to paying customers; from here on the public version line is monotonic and the upgrade path (delta packages, migration replay, in-place container swap) is fully supported.


Headline changes

1. Comms resilience — workers self-heal from cable yanks

The OPC UA and Modbus TCP pollers used to leave a stale TCP session in the per-device cache when the underlying poll was cancelled mid-call by the 10-second outer timeout. Every subsequent attempt then hung on the same dead socket until the worker process was manually restarted.

  • BaseException cleanup in both drivers — any non-clean exit (including asyncio.CancelledError from wait_for) drops the cached client / subscription session so the next cycle reconnects fresh.
  • OPC UA aliveness probe — once per poll cycle the worker reads ServerStatus.State (NodeId i=2259); a failure flips the device status to Error within seconds even when the subscription queue is silent (no value changes = no errors otherwise visible).
  • Exponential backoff per device: 5s → 10s → 20s → … → 300s cap, reset on first success. After 10 consecutive failures the device is flagged “quarantined” in the log. Earlier behaviour hammered the PLC at the configured poll interval indefinitely.
  • drop_device() hot-reload — when a device disappears from the enabled list (deleted/disabled in the UI), the worker tears down its cached state without needing a container restart.

Cable-yank end-to-end verified: 14-second recovery on OPC UA, ~40 s on Modbus (longer because pymodbus’s connect timeout is 3 s and the exponential backoff applies).

2. Telegram inbound — close the ACK loop

The platform’s Telegram channel was send-only. Operators replying with ACK-… over Telegram had their messages discarded — only SMS replies forwarded by the RUT241 webhook actually closed alarms.

  • New telegram-inbound worker — long-polls getUpdates for every active Telegram gateway. One HTTP round-trip per ~25 s when idle, near-zero quota cost.
  • Flexible matching cascade:
    1. Telegram reply gesture (swipe-to-reply) — most precise. Matches by provider_message_id.
    2. ACK-{token} — original token format still works.
    3. Simple keyword — ack, ok, okay, confirmo, confirmar, sim, yes, ✅, 👍, ✓. Matches the most recent unacked Telegram message for the sender’s chat_id within 24 h.
  • Outgoing instruction simplified — Telegram messages now read ”💬 Reply with “ack”, “ok” or 👍 to acknowledge” instead of the unfriendly ACK-pW1Wp_da-jU token. SMS / WhatsApp / email keep the precise token because their webhook parsing relies on it.
  • Confirmation reply — bot replies ”✅ Acknowledged. Escalation cancelled.” the moment a match lands so the operator knows the loop closed.

3. Migration 046 — acknowledged_by_label

alarm_events.acknowledged_by is a bigint FK to users(id) and only ever populated by in-platform UI acks. Channel-side acks (Telegram, SMS, email-token) had no way to attribute the human/contact who closed the alert in the alarm history view.

  • New column alarm_events.acknowledged_by_label TEXT populated with a readable label by every ack path:
    • "Platform: Administrator" (in-platform button)
    • "Telegram: Tiago Guimarães LA" (Telegram reply)
    • "SMS: +351912345678" (RUT241 webhook reply)
    • "Token: <contact_name>" (email magic-link ack)
  • Backfill — past in-platform acks get "Platform: <user.full_name>" from the historical users join.
  • API change — acknowledged_by_name in the alarm-history payloads is now COALESCE(acknowledged_by_label, users.full_name) so the UI shows the channel-side label without any frontend changes.

4. Migration 045 — seconds-resolution comm timeout

The “Update data timeout” was stored as integer minutes (max_minutes_without_data), too coarse for protocols with sub-second update cycles (OPC UA subscriptions, IEC 104).

  • New column devices.comm_timeout_seconds INTEGER with a check constraint of >= 5 s.
  • Backfill: comm_timeout_seconds = max_minutes_without_data * 60, defaulting NULL rows to 60 s.
  • UI shows four boxes — Days / Hours / Minutes / Seconds — and the Save handler writes the new column. The legacy minutes column stays in sync (rounded up) for any external integrations.
  • Timer is based on the most recent measurement timestamp the device reported — not on polling attempts. A subscription cycle that returns no new data does NOT reset the clock.

5. UX polish

  • Streak detection — device status flips to ❌ Error after 3 consecutive failed polls regardless of the rolling 20-poll success rate. Earlier “Intermittent” with 75 % rate was misleading after a recent cable yank.
  • Protocol-aware Intermittent hint — the popover help text now matches the protocol (Modbus / OPC UA / IEC 104 / generic) instead of always advising “increase MB_SERVER DISCONNECT”.
  • Tag Address row delete keys widgets by a stable per-row UUID so deleting a row in the middle of the list actually removes the clicked row, not the visually-last one.
  • Compact virtual-tag expression editor — Tag / Operator / Function pickers are dropdowns (scales to hundreds of tags) instead of grids. Site-scope shows a side-by-side Device + Tag pair plus the Insert button below.
  • Virtual-tag validator error humanizer — “Missing operator between ${Ambient Temp} and ${ALC105:Flow}” instead of the previous Syntax error: t76 d20t1 invalid syntax.
  • OPC UA browse lets you pick ns=0 and ns=1 (previously filtered out — useful for diagnostics on non-standard servers).
  • Reports cross-FS fix — Path.replace → shutil.move; PDF generation works under Docker where /tmp and /app/reports_output sit on different filesystems (EXDEV bug).
  • Cooldown = 0 is now valid for messaging rules. Default for new rules stays at 30 minutes. Zero = every alarm trigger sends a fresh message with no rate limit.
  • Threshold operator display — uses Unicode glyphs (≥, ≤, ›, ‹, ≠) instead of ASCII so Streamlit’s Markdown parser doesn’t swallow the leading > as a blockquote.
  • Cascade delete on tag-address Save — removing a row from the Tag Addresses tab also drops the corresponding device_tag_mappings and tag_metadata rows so the Real Time list and Variable Mapping picker stop showing the ghost name.
  • OPC UA browse and import auto-refreshes Tag Addresses + Real Time tabs without navigating away from the page.

Operator action required when upgrading from any 1.15.x

None on Postgres data — migrations 045 and 046 are additive and idempotent. Re-running them on a customer instance is a no-op.

The container image tag changes from synctide:1.15.0-test1 to synctide:1.0.1. docker compose up -d after pulling reads the new default from docker-compose.yml.