Background

The pageserver_io_operations_bytes_total metric (STORAGE_IO_SIZE) is per-timeline.

Before

• impr: propagate VirtualFile metrics via RequestContext #7202

the tenant/shard/timeline ids were inferred from the path that was passed to VirtualFile::open. The first read/write operation on that VirtualFile would instantiate the metric via with_label_values.

This happens both for Attached and Secondary locations.

Attached locations properly lifecycle-manage the metric: during Timeline::shutdown, the metric gets removed.

#7202 avoids the with_label_values in the hot path but punts on fixing the problem described below.

Problem

Secondary locations don't lifecycle-manage the metric.

Thus, if a secondary location gets detached, the instantiated per-timeline metric remains in the registry causing

• confusion
• minimal memory usage

Solution

Somehow lifecycle-manage the metric.

Secondaries don't have infrastructure for this at the timeline-level, so, some generic infrastructure needs to be put in place first.