The analytics job that filled the production disk
A change-data-capture connector streams the primary's write-ahead log into the warehouse. It crashed on Friday evening and nobody noticed, because the warehouse being stale isn't paged on.
By Sunday the primary database was out of disk and refusing every write. The replication slot outlived the consumer, and a slot holds the log until its consumer reads it — it knows nothing about whether that consumer still exists.
Boundaries, outermost first: Production database: Primary (serving writes; FAILED: disk full), Replication slot (retains the log) Outside every boundary: Warehouse, CDC connector (crashed Friday; FAILED: crashed), an empty slot for the watch the retained log Connections: Primary calls Replication slot — writes append (step 1) Replication slot publishes to CDC connector — stopped reading (step 2) CDC connector publishes to Warehouse watch the retained log controls Replication slot — retained bytes (step 3)