Ceph » bluestore

The v19.2.1 release contained a backport (https://github.com/ceph/ceph/pull/59065) with a breaking change. This PR removed the `bdev_async_discard` configuration value, and replaced it with `bdev_async_discard_thread`. This new flag is a superset of the old flag, allowing for multiple threads to be used for a task that was previously single-threaded. However, by removing the old config value from the stable v19.2.0 release breaks Ceph deployments in the following ways:

• If the value was set via ceph.conf, it is now silently ignored, and async discards are disabled.
• If the value was set via `ceph config set`, it is now silently ignored, and async discards are disabled.
• If the value is set via `ceph config set` after upgrading to v19.2.1 (via automation such as Rook), the command errors. This causes Rook deployments (new ones and upgrades) to repeatedly fail.

This may be the cause of https://tracker.ceph.com/issues/70010, as the reporter stated (on Slack) that their issue went away after they replaced the removed option with the new one.

In addition to this breaking backwards compatibility, setting the value above 1 causes all CPU cores available to each OSD to run at 100% until the value is reduced to 1 (or 0). See the below graphs for examples of this from three independent Ceph clusters and admins:

Going from unset to "4" to "1"

Going from unset to "4"

Going from unset to "4" to "2"

Related PRs:

• PR adding this to squid (contains several backports): https://github.com/ceph/ceph/pull/59065
• PR to main that removes the `bdev_async_discard` config: https://github.com/ceph/ceph/pull/55469