WAL replay in PostgreSQL

Posted May 29, 2023 20:49 UTC (Mon) by DemiMarie (subscriber, #164188)
In reply to: Cloud-storage optimizations by andresfreund
Parent article: Cloud-storage optimizations

Would using direct, async I/O with io_uring solve this problem? As long as one can queue a large number of I/O requests before needing any results, latency should not be a significant problem.

WAL replay in PostgreSQL

Posted May 30, 2023 15:12 UTC (Tue) by andresfreund (subscriber, #69562) [Link]

> Would using direct, async I/O with io_uring solve this problem? As long as one can queue a large number of I/O requests before needing any results, latency should not be a significant problem.

"It depends". Even on NVMe small random reads tend to be more expensive than larger sequential reads. On commercial clouds you pay for IOPS and also the latencies are considerably higher - making the random reads more of a problem.

We have some readahead for blocks referenced in the WAL starting in PG 15 and there's more upcoming work.