mirror of
https://github.com/neondatabase/neon.git
synced 2026-01-14 08:52:56 +00:00
95588dab98
## Problem [This select arm](https://github.com/neondatabase/neon/blob/main/safekeeper/src/send_interpreted_wal.rs#L414) runs when we want to attach a new reader to the current cursor. It checks the current position of the cursor and resets it if required. The current position of the cursor is updated in the [other select arm](https://github.com/neondatabase/neon/blob/main/safekeeper/src/send_interpreted_wal.rs#L336-L345). That runs when we get some WAL to send. Now, what happens if we want to attach two shards consecutively to the cursor? Let's say [this select arm](https://github.com/neondatabase/neon/blob/main/safekeeper/src/send_interpreted_wal.rs#L397) runs twice in a row. Let's assume cursor is currently at LSN X. First shard wants to attach at position V and the other one at W. Assume X > W > V. First shard resets the stream to position V. Second shard comes in, sees stale cursor position X and resets it to W. This means that the first shard doesn't get wal in the [V, W) range. ## Summary of changes Ultimately, this boils down to the current position not being kept in sync with the reset of the WAL stream. This patch fixes the race by updating it when resetting the WAL stream and adds a unit test repro. Closes https://github.com/neondatabase/cloud/issues/23750