🐛 Resource Timings are inaccurate when ServiceWorker is involved

[huwyca]

Describe the bug

NOTE: This was discussed and shown at length in a call with support on 12/12/2023 with traeger.meyer@datadoghq.com and shayne.sebro@datadoghq.com. There should be a recording available of that call. Datadog 1345744 // 1963 - RUM Resource Timing - Customer call

Service worker based resource network requests are being reported improperly by browser-sdk rum resourceCollection. This causes the duration to be incorrect as well as causing other fields to be missing from the RUM entry.

Browser: Chrome (Firefox and Safari don't seem to support workerStart yet) BrowserSDK versions affected: Seems to affect multiple major versions including v5 and v4

This seems to happen mainly due to requestStart being set to workerStart and those falling before the connectionStart/connectionEnd timings when a ServiceWorker. I notice it most commonly when a service worker is intercepting a request over http2 re-using an existing TCP connection to a domain.

When a service worker is involved in a network request it changes the fields of the PerformanceResourceTiming. In addition to setting the workerStart property to be non-zero it also changes the requestStart time to be set to the workerStart property. This causes a logic check to fail in the resourceUtils.ts#toValidEntry function to fail due to the entry.requestStart time falling before the other timings in the areInOrder check. https://github.com/DataDog/browser-sdk/blob/43c44563a9b6671255ee4a591197c125a33b3f68/packages/rum-core/src/domain/resource/resourceUtils.ts#L138-L157

This causes the matchRequestTiming function to not find a correct matching timing due to it filtering out the timing due to filter(toValidEntry) removing the timings from service workers. https://github.com/DataDog/browser-sdk/blob/43c44563a9b6671255ee4a591197c125a33b3f68/packages/rum-core/src/domain/resource/matchRequestTiming.ts#L24-L37

The matchingTiming being set to undefined from matchRequestTiming then results in not calculating any correspondingTimingOverrides with computePerformanceEntryMetrics. This has a couple of knock on effects:

1. Due to not having the correct entry metrics the duration being reported in the RumEvent is the timestamp of when the datadog rum XHR onLoad handler fires instead of when the resource finished. https://github.com/DataDog/browser-sdk/blob/43c44563a9b6671255ee4a591197c125a33b3f68/packages/rum-core/src/domain/resource/resourceCollection.ts#L76-L78 This makes the duration be susceptible to delays on the event loop which can cause a large difference in the reported time from the actual when the event loop is busy and a network resource is waiting to fire its onLoad handlers.

2. This also results in the various fields from computePerformanceResourceDetails not being included in the RumEvent due to not being called from computePerformanceEntryMetrics. https://github.com/DataDog/browser-sdk/blob/43c44563a9b6671255ee4a591197c125a33b3f68/packages/rum-core/src/domain/resource/resourceUtils.ts#L88-L136

To Reproduce Steps to reproduce the behavior:

I did a reproduction on a call with the support team that was recorded. Included below is a basic reproduction case but can be slightly annoying to setup and show from scratch. I'm happy to jump on a call showing a reproduction.

1. Open Chrome (firefox and safari don't have workerStart support yet) to a page with DD browser sdk running and that has a service worker registered for resources.
2. Turn on CPU throttling in the Devtools > Performance > Settings (click the gear icon on the Performance tab)
3. Take a profile while doing something that makes network requests for resources). This step is just to show that the reported duration will line up with the XHR onLoad handler time. You can view the network panel and check for the gear icon beside the network resource name for a request to see if a service worker was responsible.
4. If you log the events in the beforeSend callback you can view them easily. Otherwise you can view them in the DD rum explorer. If the workerStart is set the requestStart has been set to its value. Might take a few tries but should be able to see that the RumEvents are missing computePerformanceResourceDetails fields and have an incorrect duration.

Expected behavior RumEvents should be reported correctly when a service worker is involved. To fix this you can make a change to the toValidEntry function to check if the entry.workerStart field is non-zero. If so then use a different conditional check to verify fields are valid and in proper order.

Related useful information / links:

Spec Links:

ResourceTiming - https://w3c.github.io/resource-timing/#dom-performanceresourcetiming-requeststart NavigationTiming - https://www.w3.org/TR/navigation-timing-2/#PerformanceResourceTiming ServiceWorker - https://www.w3.org/TR/service-workers/#service-worker-timing Fetch - https://fetch.spec.whatwg.org/#http-fetch

W3C workerStart spec questions

w3c PR where they talk about how workerStart is interacting with requestStart https://github.com/w3c/resource-timing/issues/119 workerStart and redirects - https://github.com/w3c/navigation-timing/pull/131 https://github.com/whatwg/fetch/pull/1413

WPT Updates

Fetch, ResourceTiming and ServiceWorkers - https://github.com/web-platform-tests/wpt/pull/31229

[BenoitZugmeyer]

Thank you for your thorough report! We are looking into it. We'll keep you updated.

[BenoitZugmeyer]

We are still figuring how to best handle this case. In the meantime I opened a Chromium issue: https://issues.chromium.org/issues/323703325