[antiguru]

Spike, not for merge. Exposes the allocation tracker's sampled live heap as two compute introspection sources, mz_heap_live_per_operator and mz_heap_stacks, attributed to the timely operator that made each sampled allocation. The sources join with mz_dataflow_operators, so a SUBSCRIBE keeps a per-operator heap breakdown up to date. misc/spike/heap-flame.html renders one as a live flame graph over the WebSocket SQL API.

Attribution needs to know the scheduled operator at allocation time. Timely's logger cannot provide it, because its events are buffered and arrive after the allocation. The spike therefore patches timely with TimelyDataflow/timely-dataflow#831, a thread-local holding the scheduled operator's identifier. The tracker reads it only for sampled allocations. TraceManager::maintenance and allow_compaction claim the arrangement's operator, so merge work driven from compute maintenance is charged to the arrangement.

On a local 4-worker replica with lgalloc off and a 64 KiB sample interval, the tracker accounted for 2280 of 2300 MiB anonymous RSS. Operators received 92% of it, 4% was unattributed (persist blob reads on tokio threads), and the rest was the logging dataflow, mostly the in-process symbolizer. Per arrangement, sampled heap matched mz_arrangement_sizes capacity within about 1%. With lgalloc on, the sources miss lgalloc's file mappings. In a 60 s CPU profile the scheduling hook and the snapshot code did not appear, and the tracker's sampling path took about 4.6% of samples at the 64 KiB interval.

Open points before a design: release images strip DWARF, so in-process symbolization yields function names only, and stacks should be symbolized off the replica by build ID. The tag also needs a runtime component if a process hosts more than one timely runtime.

🤖 Generated with Claude Code