b2cd0da0a2
When making updates to CSI plugins, the state store methods that have open write transactions were querying the state store using the same methods used by the CSI RPC endpoint, but these method creates their own top-level read transactions. During concurrent plugin updates (as happens when a plugin job is stopped), this can cause write skew in the plugin counts. * Refactor the CSIPlugin query methods to have an implementation method that accepts a transaction, which can be called with either a read txn or a write txn. * Refactor the CSIVolume query methods to have an implementation method that accepts a transaction, which can be called with either a read txn or a write txn. * CSI volumes need to be "denormalized" with their plugins and (optionally) allocations. Read-only RPC endpoints should take a snapshot so that we can make multiple state store method calls with a consistent view. |
||
|---|---|---|
| .. | ||
| autopilot.go | ||
| autopilot_test.go | ||
| deployment_events_test.go | ||
| events.go | ||
| events_test.go | ||
| iterator.go | ||
| schema.go | ||
| schema_test.go | ||
| state_changes.go | ||
| state_store.go | ||
| state_store_oss.go | ||
| state_store_test.go | ||
| testing.go | ||