Skip to content

feat(io): batched footer resolve on rest_ioctx - #188

Draft
ran-yuan-rui wants to merge 8 commits into
NVIDIA:mainfrom
ran-yuan-rui:feat-rest-footer-resolve
Draft

feat(io): batched footer resolve on rest_ioctx#188
ran-yuan-rui wants to merge 8 commits into
NVIDIA:mainfrom
ran-yuan-rui:feat-rest-footer-resolve

Conversation

@ran-yuan-rui

@ran-yuan-rui ran-yuan-rui commented Aug 17, 2026

Copy link
Copy Markdown
Contributor

Description

Opening a parquet footer currently uses a one-shot blocking suffix GET. The process-wide CURLSH shares DNS and TLS-session state but deliberately does not share live connections, so the normal many-file path pays a separate TCP/TLS connection setup per file.

This adds rest_ioctx::resolve_footer_objects(paths, on_result, stop). A per-call curl multi submits the probes concurrently, keeps at most footer_resolve_max_inflight transfers active, and reuses idle connections within the call. Results are delivered one file at a time on the caller's thread; there is no whole-batch completion barrier. This PR adds the cuCascade mechanism only — wiring Sirius scan preparation to it follows after sirius-db/sirius#1340.

Each entry follows the existing open_io_object(path, parquet_footer_probe) behavior: verified 206 responses return the suffix window, unusable responses (200, 416, or an unverifiable 206) fall back to HEAD, a zero-byte window uses HEAD directly, and retries re-authorize each attempt while preserving ETag capture and perf-counter attribution.

Once a batch is accepted, every input occurrence is delivered exactly once. Per-entry failures (4xx, credential errors, unparsable paths) do not cancel siblings. Cancellation reports operation_canceled for undelivered entries. If a callback throws, no later entry is reported as successful; the remaining entries are canceled before the first exception is rethrown. Submission validation errors (empty batch, the API disabled, a budget below one probe window) throw before any delivery.

The returned io_object is stashless and carries only identity, size, and the validation tag. Footer bytes are returned separately under an ioctx-wide footer_resolve_stash_budget lease, released when the payload buffer is freed. Concurrent calls are FIFO-serialized per ioctx, and footer_resolve_max_inflight = 0 disables the API.

Supporting changes: exec::admission_control gains a non-blocking try_acquire plus reserved()/peak_reserved() accessors, the ioctx perf snapshot reports the budget's live/peak bytes, and the loopback test server gains keep-alive, per-key request counters, per-key response scripting, and a deterministic GET barrier.

Validation

The initial 14-case conformance suite was red against main; six cases added during review bring the final suite to 20. Coverage: byte/tag/window parity with the single-probe path including the HEAD fallback and the zero-window short-circuit, per-entry error isolation (4xx, authorizer exceptions, unparsable URIs), streamed completion order, exactly-once delivery including duplicates, cancellation before submission, mid-transfer, and during retry backoff, callback throws under simultaneously-ready completions, budget residency and gauges including sub-window rejection, single-worker and saturated consumers, FIFO fairness and queued-batch cancellation, lease-outlives-ioctx, connection reuse (12 GETs over at most 2 accepted connections in a no-retry keep-alive run), 503 retry, and snapshot parity with N single opens.

cucascade_io_tests: 135/135 green (1610 assertions).

Sirius consumer check: cuCascade c0a2607 was consumed by integrate_cucs_io at 8e7ce16b (sirius-db/sirius#1340). The full rebuild passed. make s3-test ran 92 cases: 90 passed, while the same two retry-log observability cases tracked in #175 remained failing. This PR does not change those results.

Resolving N objects' parquet footers costs N independent blocking suffix
probes, each on a fresh TCP+TLS connection (the synchronous metadata
path shares DNS/TLS-session state but not live connections).

Add rest_ioctx::resolve_footer_objects(paths, on_result, stop): batched
submission with streamed per-entry completion.  One curl multi driven on
the caller's thread carries every probe (and HEAD fallback) of a batch,
reusing its pooled connections across entries, with per-entry semantics
identical to open_io_object(path, parquet_footer_probe): verified-206
window, 200/416/unverifiable-206 HEAD fallback, the same retry policy
per entry, per-attempt re-authorization and ETag capture, and the same
perf-snapshot attribution.

Each result carries a stashless io_object (size + validation tag) plus
the footer window as a separate payload whose buffer is a lease on an
ioctx-wide byte budget: bytes return when the buffer is freed, bounding
resolve-ahead memory without attaching budget-held state to long-lived
objects.  While any transfer is active the engine acquires budget
non-blockingly; a blocking, stop-aware wait happens only at zero active
transfers.  Batches FIFO-serialize per ioctx; a queued batch cancels out
of the queue without side effects.  Delivery is exactly-once per input
occurrence on the caller's thread; cancellation aborts in-flight
transfers and delivers one operation_canceled per undelivered entry; a
throwing callback cancels the remainder and rethrows the first exception
after the sweep.

Config: footer_resolve_max_inflight (default derives n_reactors *
max_connections; 0 disables the API) and footer_resolve_stash_budget
(default 2 * inflight * footer_probe_bytes).  admission_control gains
try_acquire plus reserved/peak accessors, and the ioctx perf snapshot
reports the budget's live/peak bytes.
…gauges

The connection-reuse case retains all twelve footer payloads in its
results while sizing the stash budget at four windows; under the
payload-is-the-lease contract the resolve call then blocks waiting for
bytes only its own caller could free.  Size the budget to the retained
payload count.  The budget case now also asserts the snapshot gauges
directly (reserved == peak == budget while payloads are held; reserved
drops to zero with peak still at budget after release), and the
scheduling-loop checks use fixed assertion counts so the suite's case
and assertion totals stay stable.
Memory now tracks the ledger: every fallback, terminal, and cancel path
destroys the probe buffer before releasing its budget reservation, probe
buffers reserve exactly the window up front so growth can never exceed
the lease, and an explicit footer_resolve_stash_budget smaller than
footer_probe_bytes is rejected at submission (a sub-window budget cannot
be honored as a hard cap).

Entry submission is exception-safe: an authorizer or curl setup failure
becomes that entry's error and its siblings continue, and an unwind
guard detaches any easy handle still attached to the batch multi before
the owning entries are destroyed.  Path parsing failures likewise become
per-entry errors instead of aborting the batch with no callbacks.

The event loop drops its full-vector scans for a backoff counter plus a
deadline min-heap, giving O(N + R log N) scheduling instead of a
worst-case O(N^2), and the per-submit clock read only happens when
perf_instrumentation is on.
Three cases on the batched footer resolve: multiple malformed-206
objects under a two-window budget complete alongside healthy siblings
with the reservation gauge back at zero; an authorizer that throws for
one key delivers that exception to that entry alone, with no GET issued
for it and every sibling succeeding; and an unparsable URI in the batch
is isolated as that entry's error, while an explicit stash budget
smaller than the probe window is rejected at submission.  The loopback
harness gains per-key malformed-206 scripting and the mock authorizer
gains per-key exception injection.
After the first callback exception nothing may be delivered as success:
completion draining, new-entry admission, and retry resubmission all
stop at the first recorded throw (or a stop request), leaving undrained
completions to the cancel sweep.  A driver failure mid-batch (a curl
multi error) now runs the same sweep before rethrowing, so every
undelivered entry still receives exactly one canceled result on every
exit path.

An explicit footer_resolve_stash_budget of zero is no longer silently
treated as the derived default — it falls under the sub-window rejection
like any other too-small budget.  And a zero footer_probe_bytes now
matches the single-probe path exactly: batch entries skip the suffix
GET and start at the HEAD fallback, with no lease taken.
…nobs

A deterministic GET barrier on the loopback server holds three requests
and releases them together, so a callback that throws on the first
delivery faces simultaneously-ready siblings: the case pins one
success, every remaining entry canceled, only the initial three GETs on
the wire, the original exception rethrown, and no callback after
return.  An explicit zero stash budget throws before any authorization
or network request, falling under sub-window rejection rather than
reading as the derived default.  And a zero probe window matches the
single-probe path: no GET and no reservation per entry, one HEAD each,
with size and quoted ETag from the HEAD and a null footer.
@copy-pr-bot

copy-pr-bot Bot commented Aug 17, 2026

Copy link
Copy Markdown

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant