John Spray 651ae44569 storage controller: drop out of blocking compute notification loop if migration origin becomes unavailable (#9147) ...

## Problem

The live migration code waits forever for the compute notification hook,
on the basis that until it succeeds, the compute is probably using the
old location and we shouldn't detach it.

However, if a pageserver stops or restarts in the background, then this
original location might no longer be available, so there is no point
waiting. Waiting is also actively harmful, because it prevents other
reconciliations happening for the tenant shard, such as during an
upgrade where a stuck "drain" migration might prevent the later "fill"
migration from moving the shard back to its original location.

## Summary of changes

- Refactor the notification wait loop into a function
- Add a checks during the loop, for the origin node's cancellation token
and an explicit HTTP request to the origin node to confirm the shard is
still attached there.

Closes: https://github.com/neondatabase/neon/issues/8901

2024-10-01 07:57:22 +00:00

..

client

Remove unnecessary dependencies (#9000)

2024-09-17 17:55:45 +03:00

migrations

storcon: track preferred AZ for each tenant shard (#8937)

2024-09-06 13:11:17 +01:00

src

storage controller: drop out of blocking compute notification loop if migration origin becomes unavailable (#9147)

2024-10-01 07:57:22 +00:00

Cargo.toml

Re-export git-version from utils and remove as direct dep (#9138)

2024-09-25 14:38:35 +02:00