mirror of
https://github.com/supabase/supabase.git
synced 2026-10-10 11:55:05 +03:00
<!-- ccr-slack-attribution --> _Requested by **Ivan Vasilov** · [Slack thread](https://supabase.slack.com/archives/C063LNYJJKS/p1790710416069689?thread_ts=1790710416.069689&cid=C063LNYJJKS)_ **Before:** When a Studio API GET comes back as a 200 with an empty body, openapi-fetch hands the caller `{}` and we only see the downstream crash, with no record of the response that caused it. **After:** The first time this happens for an endpoint in a page session, Studio sends one Sentry warning, `Empty response body on successful API request`. It carries the response metadata, browser state, resource timing, and the result of a single `cache: 'no-store'` refetch. What the caller receives is unchanged. ## Problem Studio crashes trace back to GET requests that return 200 with an empty body, which openapi-fetch turns into `{}`. They are heavily skewed to Firefox and Safari. The leading hypothesis is browser cache revalidation (Express weak ETags, no `Cache-Control` on api.supabase.com), but nothing confirms it yet. The `no-store` probe tells the two cases apart: if the refetch has a body, the browser cache is the likely culprit; if it is also empty, the server or the edge is sending empty bodies. This data should show whether the fix belongs on the API side or the Cloudflare side. Context: #51041 (closed) tried to guard the crashing call sites instead. ## Needs API-side change to be fully useful Cross-origin, Studio can only read CORS-safelisted response headers, and resource timing sizes read as zero. If api.supabase.com sends `Access-Control-Expose-Headers: ETag, cf-ray, cf-cache-status, x-request-id` and `Timing-Allow-Origin: <studio origin>`, this event will also carry the ETag, cf-ray, and cache status, plus the real transfer and body sizes and the negotiated protocol. Until then, those fields read as `null` or `0`. ## Solution - `data/empty-body-diagnostics.ts` (new): `reportEmptyBodyResponse({ request, response, schemaPath })`. - Runs only for `GET` and only when `IS_PLATFORM`. Empty POST/201 bodies are legitimate. - Reports at most once per templated endpoint per page session (module-level `Set`). - Endpoint: openapi-fetch's `schemaPath` (e.g. `/platform/projects/{ref}/settings`), passed through `templateEndpointPath`. That function drops the query string and hash, replaces the segment after `projects`/`organizations`/`branches` with `{ref}`/`{slug}`/`{branch}`, and replaces UUIDs, numeric IDs, and 20+ character alphanumeric IDs with `{id}`. I used `schemaPath` rather than the request URL so user-chosen names (bucket names, function slugs) never end up in tags or fingerprints. - Probe: one plain `fetch(new Request(request, { cache: 'no-store', ... }))` with a fresh `X-Request-Id` and a 10s `AbortController` timeout. `AbortSignal.timeout` isn't available in older Safari. The probe bypasses the openapi-fetch middleware, so it can't recurse. Only the body's byte length is recorded, never its contents. - Event: `level: 'warning'`, `fingerprint: ['empty-body-response', endpoint]`, `tags: { endpoint, probe_has_body, empty_body_diagnostic: 'true' }`, where `probe_has_body` is `true` / `false` / `error`. `extra` holds: - the request: method, status, `response.type`, `redirected`, and the original `X-Request-Id` (for API log lookup) - response headers: `content-type`, `cache-control`, `last-modified`, `expires`, `content-length`, `etag`, `cf-ray`, `cf-cache-status`, `x-request-id` - browser state: `visibilityState`, `navigator.onLine`, the navigation type, ms since navigation start, and whether the page was restored from bfcache - the latest `PerformanceResourceTiming` for the URL (transfer, encoded, and decoded size, `nextHopProtocol`, `responseStatus`) - the probe: status, request ID, body length, `content-length`, `content-type`, or the error name - Fire-and-forget: everything is wrapped in a `try`/`catch`, and the caller does not await it. - `data/fetchers.ts`: the `onResponse` middleware passes `{ request, schemaPath }` to `normalizeEmptyBodyResponse`, which calls the reporter in its empty-body branch and also for a 200 that carries `Content-Length: 0`. openapi-fetch short-circuits that case to `{}` the same way, so it is the same symptom. The return value is unchanged in every branch. - `packages/common/sentry.ts`: `filterSentryEvent` normally keeps only 1% of events that aren't page crashes. It now sends events tagged `empty_body_diagnostic` unsampled, with `codeSampleRate: '1'`. A once-per-session warning would barely show up at 1%. Consent and platform gating and the third-party filter still apply. www and docs also use `filterSentryEvent`, but only Studio's reporter sets this tag, so sampling for them and for every other Studio event is unchanged. Sentry config: Studio's `beforeSend` doesn't otherwise drop this message. It has no exception values, so the no-stack-trace filter doesn't apply, and it matches no `ignoreErrors` entry. ## Review instructions 1. Check `normalizeEmptyBodyResponse` in `data/fetchers.ts`: the reporter is `void`-called and its return value is untouched. 2. Check `probe()` in `data/empty-body-diagnostics.ts`: only `byteLength` is read from the body. The probe reuses the original request's headers and credentials (same auth as the original GET). 3. Check `filterSentryEvent` in `packages/common/sentry.ts`: only the `empty_body_diagnostic` tag skips sampling. 4. Tests: `data/empty-body-diagnostics.test.ts` and `packages/common/sentry.test.ts`. ## Verification - Unit tests (`data/empty-body-diagnostics.test.ts`, new): - path templating cases - `probe_has_body` `true` / `false` / `error` - the secret body content never appears in the Sentry call - one report per endpoint - non-GET and non-platform requests are skipped - no throw when `fetch` or Sentry throws - end-to-end through `client.GET`: still resolves `{}` and reports the `schemaPath`, for both a missing `Content-Length` and `Content-Length: 0` - `packages/common/sentry.test.ts`: tagged diagnostics are sent unsampled, untagged or false-tagged ones are still sampled, and they're still dropped without consent. These tests, plus the existing `normalizeEmptyBodyResponse.test.ts`, `handleError.test.ts`, and the rest of `sentry.test.ts`, pass (67 tests) under vitest 5 + jsdom. I ran them in a minimal harness, not the full `pnpm install` workspace, because the local checkout is sparse. - I ran TypeScript 7.0.2 (`--strict`) on the five touched files against the real `api-types`, with stubbed `common`/Sentry types. No errors in the touched files. - Prettier `--check` with the repo config passes, with and without `SORT_IMPORTS=false`. - Not run locally: the full studio typecheck, `lint:ratchet`, and knip. CI covers them. The change adds no `any`, no default exports, and no deps. ## Checklist - [x] I have read [CONTRIBUTING.md](https://github.com/supabase/supabase/blob/master/CONTRIBUTING.md) 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_01UefDak8XYLMi9aiEjDPXc5 --------- Co-authored-by: Claude <noreply@anthropic.com>
73 lines
2.5 KiB
TypeScript
73 lines
2.5 KiB
TypeScript
type SentryEventTags = {
|
|
tags?: {
|
|
globalErrorBoundary?: string | number | boolean | null
|
|
third_party_code?: string | number | boolean | null
|
|
codeSampleRate?: string | number | boolean | null
|
|
empty_body_diagnostic?: string | number | boolean | null
|
|
}
|
|
}
|
|
|
|
const NON_CRASH_ERROR_SAMPLE_RATE = 0.01
|
|
|
|
export function isSentryErrorBoundaryCrash(event: SentryEventTags): boolean {
|
|
return event.tags?.globalErrorBoundary === true || event.tags?.globalErrorBoundary === 'true'
|
|
}
|
|
|
|
export function filterSentryEvent<T extends SentryEventTags>(
|
|
event: T,
|
|
{ isPlatform, hasConsent }: { isPlatform: boolean; hasConsent: boolean }
|
|
): T | null {
|
|
if (!isPlatform || !hasConsent) return null
|
|
|
|
const isErrorBoundaryCrash = isSentryErrorBoundaryCrash(event)
|
|
const isThirdPartyOnly =
|
|
event.tags?.third_party_code === true || event.tags?.third_party_code === 'true'
|
|
|
|
// Studio's once-per-session empty-body diagnostic is too rare to sample
|
|
const isUnsampled =
|
|
isErrorBoundaryCrash ||
|
|
event.tags?.empty_body_diagnostic === true ||
|
|
event.tags?.empty_body_diagnostic === 'true'
|
|
|
|
if (isThirdPartyOnly && !isErrorBoundaryCrash) return null
|
|
if (!isUnsampled && Math.random() >= NON_CRASH_ERROR_SAMPLE_RATE) return null
|
|
|
|
event.tags = {
|
|
...event.tags,
|
|
codeSampleRate: isUnsampled ? '1' : NON_CRASH_ERROR_SAMPLE_RATE.toString(),
|
|
}
|
|
|
|
return event
|
|
}
|
|
|
|
export const BROWSER_NOISE_IGNORE_ERRORS: (string | RegExp)[] = [
|
|
// === Network / infrastructure (not actionable on FE) ===
|
|
/504 Gateway Time-out/,
|
|
'Network request failed',
|
|
'Failed to fetch',
|
|
'Load failed',
|
|
'AbortError',
|
|
'TypeError: cancelled',
|
|
'TypeError: Cancelled',
|
|
|
|
// === Browser extensions & Google Translate DOM manipulation ===
|
|
'Node.insertBefore: Child to insert before is not a child of this node',
|
|
'Node.removeChild: The node to be removed is not a child of this node',
|
|
"NotFoundError: Failed to execute 'removeChild' on 'Node'",
|
|
"NotFoundError: Failed to execute 'insertBefore' on 'Node'",
|
|
'NotFoundError: The object can not be found here.',
|
|
"Cannot read properties of null (reading 'parentNode')",
|
|
"Cannot read properties of null (reading 'removeChild')",
|
|
"TypeError: can't access dead object",
|
|
/^NS_ERROR_/,
|
|
|
|
// === Non-Error throws (extensions, third-party libs throwing strings/objects) ===
|
|
'Non-Error exception captured',
|
|
'Non-Error promise rejection captured',
|
|
/^Object captured as exception with keys:/,
|
|
|
|
// === Cross-origin script errors (no useful info) ===
|
|
'Script error.',
|
|
'Script error',
|
|
]
|