Release 3.9.4

1

cluster-manager: Fixed a bug in dynamic rebalancing that could lead to repeating crashes of the cluster-manager process (process panic), volumes being left in Migrating state, a flood of repeating volume events in the event log, and increased CPU and I/O consumption by the cluster-manager. The corresponding process crashes can be identified by a panic message in journalctl ("panic: resource ... was not considered as under update, task ID: ..., operation: PgMigrationOperation, state: Running, resource: ...").

LBM1-34103

2

cluster-manager: When a migration task is failed or aborted during dynamic rebalancing, the corresponding etcd updates are now performed as one atomic transaction to prevent partial updates. This fix prevents race conditions that could lead to cluster instability, resulting in the inability to perform management operations or to properly access volumes.

LBM1-33412

3

lightbox-exporter: Fixed a server performance dashboard exporter port problem that caused empty selector and "no data" in “FE General Statistics”, “FE Write Statistics” and “FE Read Statistics” graphs.

LBM1-34171

4

monitoring stack: Fixed the server performance dashboard "exporter port"issue by adding back the default exporter port 8090 and default 1m refresh.

LBM1-34203

5

monitoring stack: Fixed the server performance dashboard read throughput panel, showing all nodes in the cluster.

LBM1-35011

6

node-manager: Increased node-manager.service shutdownCompleteTimeout and the internal shutdownCompleteTimeout and recoveryCompleteTimeout to be sufficiently long for systems with large storage size. This prevents timing out during a graceful shutdown (resulting in an unwanted abrupt powerup), or during an abrupt powerup (resulting in failure to power up the node). Note that this change is only applicable to new installations.

LBM1-34417

7

packages/repository: Fix printing of etcd transactions in logs.

LBM1-34077

8

userlbe: Implemented hysteresis for metadata (MD) ranges and storage over provisioned space (OVP) read-only/read-write flickering. Under certain conditions - e.g., running out of MD shards and/or running out of storage OVP - it would go into read-only, and then immediately go back to read-write when going below the threshold again. This could lead to quick flickering between read-only and read-write states, which is undesirable to filesystems. Implement a hysteresis mechanism to increase the likelihood that when going into read-only or back to read-write, it will stay for a longer period of time.

LBM1-34172

9

userlbe: Sped up graceful shutdown by removing a (no longer needed) limit of only using four cores.

LBM1-34180