How long your data is kept

The table

What happens at the boundary

While a reading is inside its full-resolution window, it exists exactly as you sent it — every individual sample, at whatever rate you wrote it.

When it passes that window the individual reading is deleted and the hourly summary covering it remains: minimum, maximum, average and count for that tag for that hour. Nothing is parked in an offline archive you have to request and wait for — a query that outruns the raw window is answered from the aggregate tier automatically, and it returns numbers rather than an empty chart.

What that means in practice

• “Was the freezer warmer last July than this July?” — answered on every plan, including free. That is an hourly-average question.
• “What was the exact pressure trace during the trip at 03:14 last Tuesday?” — needs the raw window. 14 days on Free, 1 year on Pro, 2 years on Business.
• “Did the peak ever exceed the limit in 2027?” — answered from the hourly maximum, years later, on any plan. A spike that lasted a second still shows up in that hour's maximum even after the individual readings are gone.

Downgrades and failed payments

A downgrade never destroys history. Lowering a tier — voluntarily or through a failed payment — stamps a 90-day retention floor during which the old tier's retention still applies. That is a full quarter to notice, export, or pay back in before anything shorter starts deleting.

The honest limit of that promise: the floor is stamped on a tier decrease. A customer who has never changed tier has never had one, so it is protection against a downgrade rather than a general grace period.

Getting your data out

Export everything, in Settings in the Explorer, gives you your account's measurements, tags and settings in one zip archive:

• every tag's readings still kept at full resolution — your plan's window — one file per month, each value exactly as stored;
• its hourly and daily minimum, maximum, average and count over its whole history, up to 10 years;
• every tag's settings, labels and former names, your alert rules, notification channels (without their tokens or keys) and chart views, and your account record;
• a README that explains every file and every column, so the archive makes sense without us.

The alerts your rules raised and the account's users are not in the archive. The API returns them as JSON (GET /api/alerts/history and GET /api/team/members); section 9 of the privacy policy lists every such call.

Readings come as CSV or Parquet, your choice. The archive is built in the background, at a pace that leaves incoming data its share of the service — minutes for most accounts, hours for the largest, and longer if another account's export is being built first — and we email you when it is ready. Settings shows how far it has got. It can be downloaded for 7 days, and you can make one every 24 hours. Through the API it is POST /api/export/everything.

A tag at a time, measurements, aggregates and tag metadata also export straight from the API as Parquet, CSV or JSON, with no wait. No export fee either way, no proprietary step in the middle.

For a report to send on rather than a copy of everything, an Excel report gives the tags and period you pick a row per hour or per day, with minimum, maximum and average, from the hourly summaries — so it reaches as far back as they do. On Pro and Business.

One clarification, because this site got it wrong once and it is worth being precise: Parquet is an export format. It is not how your data is stored internally — internally it is a compact fixed-size binary record format of our own. If a page ever tells you your aged data is “stored as Parquet”, that page is wrong.

Long ranges are not slow

A five-year chart is not five years of raw points sent to your browser. The server reduces the series before it leaves — maxPoints with decimation=lttb keeps the shape of the curve, so spikes, steps and outliers survive rather than being averaged away by every-Nth-sample thinning. A five-year range is answered out of the day tier at roughly 1 800 records instead of hundreds of millions of raw samples.

An hourly or daily point is an average, so it never travels alone: each one carries the lowest and highest value of its hour or day as minValue and maxValue, and when the reduction drops an hour, the point drawn nearest to it takes over its extremes. A two-minute spike is still there on a two-year chart; Trends draws it as a band around the line.

A chart whose resolution is picked for it - Trends on Auto, a shared link, a maxPoints request - reaches the present at hourly and daily resolution too. An hour is summarised about ten minutes after it ends, so until then such a chart works its newest hours - and today, on a daily chart - out of the readings themselves, the same way the summary will, and says so: each such point carries provisional: true. The hour in progress, and today, go on changing until they are over. What reads the summaries alone ends at the newest one: a fixed “1 hour average” or “1 day average” of an analog tag in Trends over more than two days, and a report you download. A CSV saved from Trends leaves out the hours not summarised yet.

maxPoints is clamped to 10 000, and asking for decimation without it is a 400 rather than a guess about what you meant.

Where the data physically is

Azure Sweden Central, with a nightly backup to a geo-redundant vault whose offsite copy sits in the paired Swedish region. The details, including both recovery-point objectives and which restore paths have been rehearsed and which have not, are on where your data lives.