[antiguru]

Compute rendering wraps imports, objects, let bindings, the main body, and joins, reduces, top-k and sinks in timely regions. Each region is a full subgraph with its own progress tracker and all-worker progress exchange, so a frontier advance is propagated and broadcast once per region it crosses, on every worker. For replicas with many idle dataflows this per-tick bookkeeping dominates CPU, and it grows with the worker count rather than with data.

This change renders those regions inline: the closure runs in the enclosing scope and entering or leaving is the identity, through crate-local inline_named, enter_inline and leave_inline in render/inline_region.rs. Operators are unchanged. BuildRegion stays a real region, because index imports enter it from the dataflow scope.

Measured locally on 16 workers with static tables ticking once per second, upstream/main against this branch: 48 dataflows go from 1.38 to 1.00 cores (-28%), 426 dataflows from 12.5 to 8.4 cores (-32%), with scopes dropping from 7,968 to 2,252, reachability log records by 82%, and hydration from about 65 s to 50 s. Lag is unchanged.

Introspection gets shallower: mz_dataflow_addresses and mz_dataflow_operator_parents lose the region levels, the dataflow visualizer nests less, and mz_lir_mapping operator ranges cover fewer operators. introspection/relations.slt changes accordingly. Testdrive files that inspect operator addresses may need updates once CI runs them.

🤖 Generated with Claude Code