Root causes are fixed: Realtime now purges both projects:$projectId
and users:$userId from the shared cache before the respective reads,
eliminating the cross-process negative-cache races that surfaced as
intermittent 'Session is not valid.' failures on
appwrite-labs/cloud#3214's Realtime (dedicated) E2E.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Cross-process negative-cache race in Realtime's onMessage handler: when
the WebSocket opens against a just-created project, an earlier router
probe or subscription lookup may have cached projects:\$projectId as
empty before the project actually existed. Although the HTTP worker
purges on createDocument, a narrow window lets Realtime re-populate the
stale empty entry in its own process cache, after which getProjectDB()
sees an empty Document and falls back to getConsoleDB(), which looks
for users in the console namespace and returns nothing — sessionVerify
then fails with 'Session is not valid.'
Surfaces on cloud#3214 as a recurring Realtime (dedicated) E2E failure.
Diagnosed via [realtime-project-diag] + [realtime-auth-diag]
instrumentation — the logs show consoleDb=appwrite consoleNs=console15x
(correct), project lookup empty, and subsequent user lookup landing in
the console namespace.
Purge is bounded: one cache DEL per WS message (not per read), runs
only when a projectId is present, and only forces the single
projects:\$projectId key to be re-read from adapter.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
When a DocumentsDB collection's metadata document exists but the
backing store (e.g. a freshly restored dedicated shard) has no table,
Utopia\Database\Database::find() throws NotFound. Previously this
bubbled to the HTTP layer as a generic 500. Catch it in the same
try/catch that already handles OrderException, QueryException, and
Timeout, and translate into COLLECTION_NOT_FOUND/TABLE_NOT_FOUND via
getParentNotFoundException() so the caller sees the correct 404.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Temporary instrumentation to diagnose the 'Session is not valid.' failure
on appwrite-labs/cloud#3214's Realtime (dedicated) E2E. Emits:
- project ID in scope
- user ID from the session payload
- whether the user document was found in the project DB
- session count on the returned user document
- session-secret prefix being verified
- sessionVerify result
This will reveal whether the race is (a) user doc missing in project DB
(routing/provisioning issue) or (b) user doc found but sessions array
empty or mismatched (session write path issue). Revert once the root
cause is identified.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Cross-process read-after-write race: the HTTP worker writes a new session
into the user document on /account login endpoints, then the client sends
an authentication frame over a different Swoole process (Realtime). Cache
propagation between processes is not guaranteed to be observed on the
very next read, so sessionVerify() occasionally fails with a stale user
document whose sessions array does not yet contain the just-created one.
Purge the cached user locally before the read so sessionVerify sees the
freshly-written session deterministically. Overhead is bounded: one DEL
on cache + one extra primary-key read, executed once per WebSocket
authentication frame (not per message).
Surfaces on PR appwrite-labs/cloud#3214 as an intermittent
'Session is not valid.' failure in RealtimeConsoleClientTest +
RealtimeCustomClientTest manual-authentication cases in dedicated mode.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add ProviderRepositoryFrameworkList and ProviderRepositoryRuntimeList
model classes with conditions and type field so the listRepositories
endpoint's oneOf response gets a discriminator on the type property.
Each Detection subclass now declares only its own type value in the enum
rather than sharing the full ['runtime', 'framework'] list. This prevents
SDK validators from accepting invalid values on concrete models.
Unify getDiscriminator to produce a single discriminator object for both
single-key and compound cases. Single-key returns standard {propertyName,
mapping}. Compound falls back to extending the object with x-propertyNames
and x-mapping for multi-property discrimination.
Simplify call sites: OpenAPI3 uses 'discriminator', Swagger2 uses
'x-discriminator' — no more split keys.
Add conditions to all 7 Algo models (AlgoArgon2, AlgoBcrypt, AlgoMd5,
AlgoPhpass, AlgoScrypt, AlgoScryptModified, AlgoSha) to enable
discriminator generation for hashOptions unions.
Fix misspelled method name (getDisciminator -> getDiscriminator) across
Format, OpenAPI3, and Swagger2. Extract duplicated model-resolution
lambda into Format::resolveModels(). Fix copy-pasted descriptions in
ProviderRepository list models.
The SDK push task used `git push --force-with-lease` which fails on
repos with branch protection rules that disallow force pushes. Instead,
checkout the existing remote dev branch and commit on top of it so a
regular push is always a fast-forward.
testEmailPasswordSessionNotCorruptedByConcurrentRequests relies on
timing-sensitive curl_multi orchestration with hardcoded delays to
reproduce a cache race window. This makes it inherently flaky in CI
where resource pressure shifts the timing unpredictably.
ProjectWebhooks tests have shared state dependencies (e.g. index
creation must complete before assertions). Running with --functional
(parallel methods) causes flaky failures where indexes are still
'processing' instead of 'available'.
Jake BarnbyandClaude Opus 4.7
Chirag AggarwalandGitHub
Atharva Deosthale
Damodar LohaniandGitHub
Damodar Lohani
Matej BačoandGitHub
Matej Bačo
ArnabChatterjee20kandGitHub
ArnabChatterjee20k