browser-db-sqlite-persistence: closing the leader tab fail-stops follower collections with IndeterminateCommitError #2085
Send to my agentSummary#
With BrowserCollectionCoordinator, closing the leader tab often puts collections in the remaining tab into a permanent error state. The follower's in-flight rpc:applyCommittedTx for a sync-sourced transaction (an Electric change) waits for the closed leader, times out after 10 s, crosses leadership, and rejects with IndeterminateCommitError. The collection fail-stops, and the tab stops getting live rows.
Versions#
@tanstack/db0.12.3@tanstack/db-sqlite-persistence-core0.4.7@tanstack/browser-db-sqlite-persistence0.2.32 (OPFS,openBrowserWASQLiteOPFSDatabase)@tanstack/electric-db-collection0.5.8- Chromium (Playwright headless shell), two tabs in one browser context, same origin
Setup#
- One OPFS database, shared by the tabs through
BrowserCollectionCoordinator({ dbName })andcreateBrowserWASQLitePersistence({ database, coordinator }). - Seven persisted Electric collections (eager and on-demand), created with
persistedCollectionOptions({ ...electricCollectionOptions(...), persistence, schemaVersion: 1 }).
Steps#
- Open tab A. It becomes the leader for every collection.
- Open tab B. Wait until both tabs show live rows.
- Close tab A (
page.close(), a normal close, not a crash). - Keep inserting rows on the server, so that tab B's Electric streams deliver changes.
Observed#
3 of 6 runs. About 10 s after tab A closes, tab B logs this (ids shortened):
IndeterminateCommitError: Commit outcome is indeterminate for collection "messages":
rpc:applyCommittedTx:req crossed leadership from 65c2… (term 1) to 2f07… (term 2)
<- Error: RPC rpc:applyCommittedTx:req timed out after 10000ms
at BrowserCollectionCoordinator.assertMutationRouteUnchanged
at BrowserCollectionCoordinator.sendRPC
at async PersistedCollectionRuntime.persistAndBroadcastExternalSyncTransactionUnsafe
at async PersistedCollectionRuntime.applyBufferedSyncTransactionUnsafe
The same error occurred for other collections in other runs, sometimes through flushQueuedHydrationTransactionsUnsafe. The collection stays in error, and every later update() on it throws. While tab A closes, it logs PersistedCollectionDurabilityError ... The page is closing its SQLite connection. In the other 3 runs, tab B took over every leader lock, and sending and receiving continued.
Expected#
A normal leader-tab close should not fail-stop the other tab's collections. #1845 makes a leader change during a mutating RPC indeterminate and leaves reconciliation to the application. For a transaction that comes from the sync source, the source can deliver the same changes again from the durable resume offset. So one of these seems possible:
- The closing leader rejects or drains pending RPCs (for example on
pagehide), so the follower does not wait 10 s for a leader that has gone. - For sync-sourced transactions, the runtime reconciles by itself: it re-hydrates and resumes from the persisted offset instead of fail-stopping.
- If the application must reconcile, document how to do it for a sync-sourced collection. Is
collection.cleanup()followed by a new collection the intended path?
Not done#
- No standalone reproduction yet. We saw this in an app, driven by Playwright. We can try to reduce it if that helps.
Sign in to join this conversation.