history-paging.md

History continuation paging measurements

The continuation_pages_complete_long_changing_histories_without_cursor_walk test completes every page of 201- and 302-commit changing histories, with new readers and sessions for each request, in owner/public views and denial OFF/ON. Canonical bytes and IDs match the fixture's commits in first-parent order. Every page performs exactly two raw range GETs per returned commit, with zero history/tree cursor walk. The independent continuation_selected_ref_overhead_is_bounded_independently_of_cursor_depth test measures six additional physical calls/rounds with denial OFF and seven with denial ON, in both views, relative to the selected-ref history primitive.

Reproduce from rust/:

CARGO_PROFILE_DEV_DEBUG=0 cargo test --locked -p mkit-server --all-features --lib continuation_pages_complete_long_changing_histories_without_cursor_walk -- --ignored --nocapture --test-threads=1

The default lane covers a 40-commit changing history over four/six pages, using the same canonical-byte, budget, fixed-expiry and no-cursor-walk assertions in all view/denial/latency configurations. The full fixture runs explicitly in the serial ignored-lane profile in GitHub Actions and Cloud Build. Its standalone debug measurement used 13.10 s CPU and 14.17 s wall time; parallel Cloud Build exceeded the default 60 s ceiling. The slow lane retains every assertion and a bounded 300 s timeout. No production allowance or general timeout changes.

These are bounded memory-store dispatch measurements with a fixed business clock. Reader units retain the 8,500-unit request allowance, including parsing, security reads and page work. Reader construction is included in physical calls; its separate authorization allowance is outside the session ledger. The 201-commit fixture has one real indexed pack per commit and distinct nested snapshots. The 302-commit fixture packs two commits per upload, with distinct files and trees, to remain within the existing fixture write quota. No reader or production write allowance is increased. The first paging request in a repository may perform one coordinator apply to activate its retained visibility revision; following visibility writes maintain that fence even without token issuance configured.

Physical calls are KV dispatches plus twice the ranged GET count, a conservative model covering stores that check metadata before fetching a range. The read ledger also charges two units. Paused Tokio time injects 30 ms and 130 ms RPC waits through the reader-batching latency probe. All continuation reads share the existing six-call I/O admission envelope. Blob delays include two sequential range RPCs. The fixture performs reader construction and guarded applies without delay; the reported wait adds their physical calls serially at the selected RPC latency. The table records virtual wait and its equivalent RPC rounds, including Tokio timer granularity. It does not assume a fixed number of denial waves.

The business clock stays fixed while virtual time measures storage waits. These figures are recorded without a latency gate and are not observed deployment latency or p95. Real expiry, deadlines, queues, CPU and nonempty denial directories can terminate slower requests. The measurements establish paging budget completion. Canonical windows and permission/denial caches are absent.

Timestamp-discovery continuation pages

The timestamp_pages_match_reducer_and_stay_bounded test pages a 52-commit linear history and a 52-commit merge-heavy history in canonical timestamp order (page size 10): the embedder drives page 1 with per-commit canonical reads and issues a version-2 continuation, then fresh readers and sessions redeem server pages until the walk completes. Concatenated IDs and canonical bytes match an in-memory TimestampDiscovery replay of the fixture graph. Pending keys are either all hydrated before issuance (hydrate=true) or left unknown for the redemption to hydrate in batches (hydrate=false). Each measured page is one fresh request with the reader-batching latency probe; blob loads equal emitted plus hydrated objects and no page performs a cursor walk, a ref-index or all-ref partition scan, or a reader_tips call. All rounds are far below the 400-round page budget.

Reproduce from rust/:

CARGO_PROFILE_DEV_DEBUG=0 cargo test --locked -p mkit-server --all-features --lib timestamp_pages_match_reducer_and_stay_bounded -- --nocapture