mirror of
https://github.com/neondatabase/neon.git
synced 2026-05-20 06:30:43 +00:00
cb8782a51a
Previously, it was possible for backends to request a page with the LSN of the record currently being replayed. This could cause a deadlock when the redo process wanted to read that same page at the same time. This LSN could only appear when the page was not present in the LwLSN cache, and the highest evicted LSN also was the LSN of the currently-replayed WAL record. The issue is fixed by splitting maxLastWrittenLsn into two: one for data pages, and one for metadata. This allows us to keep track of metadata changes separately, removing the implicit dependency of page IO on metadata LSNs where appropriate. Additionally, we stop evicting LwLSNs for pages with an LSN that is yet to be replayed. This means the global data page LwLsn will never return an LSN of a record that has yet to be replayed, *unless* the startup process has already determined that it won't access that page again, making page IO and Replay waits by other backends using that LSN safe for those pages.