John Spray 887e94d7da page_service: more efficient page_service -> shard lookup (#6037) ...

## Problem

In #5980 the page service connection handler gets a simple piece of
logic for finding the right Timeline: at connection time, it picks an
arbitrary Timeline, and then when handling individual page requests it
checks if the original timeline is the correct shard, and if not looks
one up.

This is pretty slow in the case where we have to go look up the other
timeline, because we take the big tenants manager lock.

## Summary of changes

- Add a `shard_timelines` map of ShardIndex to Timeline on the page
service connection handler
- When looking up a Timeline for a particular ShardIndex, consult
`shard_timelines` to avoid hitting the TenantsManager unless we really
need to.
- Re-work the CancellationToken handling, because the handler now holds
gateguards on multiple timelines, and so must respect cancellation of
_any_ timeline it has in its cache, not just the timeline related to the
request it is currently servicing.

---------

Co-authored-by: Vlad Lazar <vlad@neon.tech>

2024-01-16 09:39:19 +00:00

..

benches

pageserver: make walredo shard-aware

2024-01-03 14:22:40 +00:00

client

pageserver: API updates for sharding (#6330)

2024-01-16 09:21:00 +00:00

ctl

pagectl: add subcommand to rewrite layer file summary (#5933)

2023-11-30 11:34:30 +00:00

pagebench

fix(pagebench): #6325 broke running without --runtime (#6351)

2024-01-15 08:54:19 +00:00

src

page_service: more efficient page_service -> shard lookup (#6037)

2024-01-16 09:39:19 +00:00

test_data

Add walingest test (#5892)

2023-11-28 12:50:53 -05:00

Cargo.toml

add a Rust client for Pageserver page_service (#6128)

2023-12-18 18:17:19 +00:00