Performance
How Sypnia measures page speed and API speed for real people, how percentiles are computed and shown, and how it names the element behind a slow page.
This page is for anyone who wants to know which pages and API calls are slow for real people, and how far to trust each number. The measurements and the numbers are part of the open-source engine. Sypnia Cloud adds the performance pages of the app; on the engine you read the same numbers through the API, the CLI and MCP.
All numbers come from real visits to your app, not from a test run on one machine.
Page speed
For every page a person opens, the browser SDK reports four measurements.
| Measurement | What it tells you | Good | Poor |
|---|---|---|---|
| LCP, largest contentful paint | How long until the main content was visible. | up to 2,500 ms | above 4,000 ms |
| INP, interaction to next paint | How long the page took to react to a click, a tap or a key. | up to 200 ms | above 500 ms |
| CLS, cumulative layout shift | How much the page jumped around while the person looked at it. A score, not a time. | up to 0.1 | above 0.25 |
| TTFB, time to first byte | How long your server took to start answering. | up to 800 ms | above 1,800 ms |
Between «good» and «poor» a value is rated «needs improvement». These are the public Web Vitals thresholds.
Rules of measurement:
- LCP and TTFB belong to the page where the browser loaded the document. Later pages reached by in-app navigation have none.
- INP and CLS are measured for each page, including pages reached by in-app navigation. A page’s INP is its slowest interaction. A page’s CLS is the sum of its shifts.
- Values that cannot be real are dropped: an LCP or TTFB above 120 seconds, an INP above 60 seconds, a CLS above 10.
The element behind the number
A slow LCP is only useful if you know what was slow. Each measurement names its element:
| Measurement | Element | Extra |
|---|---|---|
| LCP | The element that was the main content. | Whether it is an image or text. |
| INP | The element the person interacted with. | Whether it was a pointer or the keyboard. |
| CLS | The element that moved the most. | |
| TTFB | None. It is about the server. |
The element is named by a selector built from its test id, its id or a short path of tags. The selector never contains the page’s text. For each page, Sypnia lists the elements that most often stand behind each measurement, with their own numbers.
API speed and failures
For every call the page makes to your API, and every request your server handles, Sypnia keeps numbers for each method and route:
| Number | Meaning |
|---|---|
| Calls | How many. |
| p50 and p95 duration | As the browser saw it, network included. |
| Server p95 | As the server measured it, when a server SDK or OpenTelemetry reports the request. |
| Server errors and no answer | Calls that answered 500 or above, or got no answer at all. |
| Rejected | Answers from 400 to 499. Usually your app saying no on purpose. A 400 can also open a problem; see Problems. |
| Failure rate | Calls that failed, divided by calls. A call has failed when it got no answer, answered 400, or answered 500 or above. Other answers from 401 to 499 are rejected calls, not failures. |
| Calls behind it | The average number of calls each request makes to a database or another service, and the average time it spends in them. |
Routes are listed slowest first, by p95. A route is listed as failing when more than 2 % of its calls answered 500 or above, or nothing, with at least 50 calls in the period. That is also the rule of the «API failing» alert, which looks at the last hour.
How percentiles are computed and shown
Averages hide slow visits, so Sypnia shows percentiles.
- p75 is the headline for pages: three visits out of four were this fast or faster. A page’s rating (good, needs improvement, poor) is the rating of its p75.
- p50 and p95 are shown for API calls: the typical call, and the slow tail.
Percentiles are computed from histograms with fixed buckets, so any range of hours or days can be added together and still give a percentile. The buckets are:
| For | Bucket bounds |
|---|---|
| Durations, in ms | 50, 100, 200, 300, 500, 750, 1,000, 1,500, 2,000, 3,000, 5,000, 8,000, 12,000, 20,000 |
| CLS | 0.01, 0.025, 0.05, 0.1, 0.15, 0.25, 0.4, 0.6, 1 |
This has three consequences, and Sypnia shows each of them instead of hiding it:
- A percentile is exact to its bucket, not to the millisecond. Inside a bucket the value is estimated in proportion to the samples. A p75 of 3,100 ms means «between 3,000 and 5,000 ms, close to the lower end».
- Above the last bound, the real value is unknown. A percentile that lands beyond 20,000 ms, or beyond a CLS of 1, is shown as «more than 20 s», and the data carries a flag that says so.
- A percentile comes with its number of samples. A p75 from 12 visits and one from 12,000 are not worth the same, so the count is always next to the number. Lists of the slowest pages, and of pages that got slower, leave out pages with fewer than 50 views in the period.
Numbers can be read for one device class: mobile, tablet or desktop. They are complete within 10 minutes.
When a page «got slower»
Sypnia says a page got slower when both are true:
- its p75 LCP over the last 24 hours is 30 % or more above its p75 over the 7 days before, and
- the page had at least 50 views in each of the two periods.
When that happens Sypnia creates an alert, once, and the daily summary reports it. Both numbers can be changed in the project’s alert rules. See Alerts and the daily summary.
Weekly suggestions
Once a week, on Monday at 07:00 UTC, the agent reads the numbers of the week that ended, the elements behind them and the click data, and proposes improvements to performance and to how the app is used, each with its evidence: the page, the number, the element. Some suggestions come from fixed rules, such as an INP above 500 ms at p75 or more than 3 rage clicks for each 100 views, and need no model. The others are written by a model. See How the analysis works.
| Sypnia engineOpen source, self-hosted | Sypnia CloudPaid service, with a free tier | |
|---|---|---|
| Measurements, percentiles, elements | Yes. Read them through the API, the CLI or MCP. | Yes, with the performance pages of the app. |
| Weekly suggestions | Those from fixed rules, always. Those written by a model, with your own model key. | Included. |
| When a page gets slower | An alert is recorded and sent to your webhooks. | An alert by e-mail and in the app, and a line in the daily summary. |
Example: the search page
The shop’s daily summary says: «The search page got slower: 3.1 seconds to load, it was 1.9.»
Behind that sentence:
| Previous 7 days | Yesterday | |
|---|---|---|
| Route | /search | /search |
| p75 LCP | 1,900 ms (good) | 3,100 ms (needs improvement) |
| Views | 1,804 | 262 |
3,100 is 63 % above 1,900, and both periods have more than 50 views, so the page «got slower». The element behind the LCP is the first product image, an image, in most of yesterday’s visits. That points at the picture, not at the server: TTFB did not change.
Reference
The fields and filters are in the API reference. What the browser SDK measures and how to turn it off is in Browser.