merlimat opened a new pull request, #25960:
URL: https://github.com/apache/pulsar/pull/25960
### Motivation
`MetadataTransactionBuffer` kept an in-memory entry for **every**
transaction it ever saw (committed and aborted alike) in its `txns` cache, and
never removed them — so the map grew unbounded for the segment's lifetime. On
long-lived segments with high transaction throughput this is a memory leak.
### Modifications
- `applyTerminalNow`: drop the entry from `txns` once the transaction is
terminal, bounding the cache to the current open-transaction count. This is
safe:
- `recomputeMaxReadPositionLocked` only consults `OPEN` entries, so
dropping terminal ones doesn't affect the max-read position;
- `isTxnAborted` reads the **separate** `abortedTxns` set — an aborted txn
stays filtered there, and a committed/unknown txn correctly reads as visible;
- the positions needed for the durable side-effects (aborted record,
watermark, op-record cleanup) are captured before the entry is removed.
- Add `trackedTxnCount()` (`@VisibleForTesting`) and a unit test asserting
the cache is pruned back to empty after a mix of commits and aborts, while
visibility stays correct (aborted filtered, committed visible).
`MetadataPendingAckStore.openTxns` is already pruned on terminal, and `/txn`
headers are already GC'd by the coordinator's GC sweep, so this change targets
the one remaining unbounded structure.
### Out of scope (follow-up)
The durable per-segment aborted records and the `abortedTxns` set grow only
with **aborts** and are meant to be pruned on segment-ledger trim. That
trim-driven GC needs a ledger-trim trigger that doesn't exist yet
(`TransactionBuffer.purgeTxns` is declared but never invoked anywhere — the
legacy buffer no-ops it too), so it's left as a follow-up.
### Verifying this change
- `MetadataTransactionBufferTest` — 13/13 (12 existing + the new
`terminalTxns_prunedFromCache_visibilityUnchanged`).
- `V5TransactionTest` end-to-end commit/abort — passing.
- checkstyle clean.
### Does this pull request potentially affect one of the following parts:
- Dependencies: no
- The public API: no
- The schema: no
- The default values of configurations: no
- The threading model: no
- The binary protocol: no
- The REST endpoints: no
- The admin CLI options: no
- Anything that affects deployment: no
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
