Question: RPO may not equal to 0?
In PageServer, data is layered and only aggregated to a certain extent before being uploaded to cloud storage. If the PageServer in the public cloud uses EKS + instance storage, does this mean there’s a chance of RPO > 0 when the PageServer crashes?
ref to: https://github.com/neondatabase/neon/blob/main/docs/pageserver-storage.md
GitHub
neon/docs/pageserver-storage.md at main · neondatabase/neon
Neon: Serverless Postgres. We separated storage and compute to offer autoscaling, code-like database branching, and scale to zero. - neondatabase/neon
6 Replies
flat-fuchsia•3mo ago
Safekeeper should ensure that WAL lives long enough for pageserver to save layer files to S3
There's also the concept of secondary pageservers. Those can be used instead if one pageserver fails and needs to recover WAL before being available again.
other-emeraldOP•3mo ago
is there a "hard" way that safekeeper ensures that WAL lives longer than pageserver needs? e.g.: pageserver may issue a TRUNCATE command to safekeeper?
flat-fuchsia•3mo ago
other-emeraldOP•3mo ago
Thank you. you mean safekeeper will dump WAL asyncly to S3? 3 copies or 1? if 1, how do safekeepers know who the "leader" is?
flat-fuchsia•3mo ago
That I don't know, sorry
other-emeraldOP•3mo ago
Thank you for your reply:)