Add io-uring based ObjectStore for local file I/O#21673
Closed
Dandandan wants to merge 9 commits intoapache:mainfrom
Closed
Add io-uring based ObjectStore for local file I/O#21673Dandandan wants to merge 9 commits intoapache:mainfrom
Dandandan wants to merge 9 commits intoapache:mainfrom
Conversation
Introduces `datafusion-object-store-iouring`, a new crate that provides an `IoUringObjectStore` using Linux's io_uring interface for high-performance local file reads. A dedicated thread runs an io_uring event loop, and read requests (`get_opts`, `get_ranges`) are dispatched via channels — enabling batched syscalls where multiple byte-range reads (e.g., Parquet column chunks) are submitted in a single `io_uring_enter()` call instead of individual `pread()` calls. Key design: - Dedicated io_uring worker thread with a 256-entry submission queue - Unbounded mpsc channel for requests, oneshot channels for responses - Range reads batched per-request; chunked if exceeding ring capacity - Write/list/copy/delete operations delegated to LocalFileSystem - On non-Linux platforms, all operations fall back to LocalFileSystem - Feature flag `io-uring` on `datafusion-execution` to opt in Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Resolve merge conflicts with upstream (version 53.0.0, object_store 0.13.2) - Update ObjectStore trait impl for 0.13.2 API changes: - copy/copy_if_not_exists → copy_opts(CopyOptions) - delete → delete_stream (required method) - PutMultipartOpts → PutMultipartOptions - Import ObjectStoreExt for head() convenience method - Enable io-uring feature by default in datafusion-execution Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Contributor
Author
|
run benchmarks |
|
🤖 Benchmark running (GKE) | trigger CPU Details (lscpu)Comparing worktree-io-uring-object-store (0bf32dd) to 7bfa3fb (merge-base) diff using: tpcds File an issue against this benchmark runner |
|
🤖 Benchmark running (GKE) | trigger CPU Details (lscpu)Comparing worktree-io-uring-object-store (0bf32dd) to 7bfa3fb (merge-base) diff using: tpch File an issue against this benchmark runner |
|
🤖 Benchmark running (GKE) | trigger CPU Details (lscpu)Comparing worktree-io-uring-object-store (0bf32dd) to 7bfa3fb (merge-base) diff using: clickbench_partitioned File an issue against this benchmark runner |
|
🤖 Benchmark completed (GKE) | trigger Instance: CPU Details (lscpu)Details
Resource Usagetpcds — base (merge-base)
tpcds — branch
File an issue against this benchmark runner |
|
🤖 Benchmark completed (GKE) | trigger Instance: CPU Details (lscpu)Details
Resource Usageclickbench_partitioned — base (merge-base)
clickbench_partitioned — branch
File an issue against this benchmark runner |
- Fix missing `StreamExt` import for `.boxed()` on Linux code path - Fix `GetResult.range` type: `Range<u64>` in object_store 0.13.2 - Add `io-uring` feature to datafusion core, forwarding to execution - Add to core's default features so benchmarks get it automatically Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
CI runners and Docker containers often block io_uring_setup via seccomp filters (EPERM). Instead of failing hard, probe availability at construction time and gracefully fall back to LocalFileSystem for all read operations when io_uring cannot be initialized. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Contributor
Author
|
run benchmarks |
|
🤖 Benchmark running (GKE) | trigger CPU Details (lscpu)Comparing worktree-io-uring-object-store (87539f6) to 7bfa3fb (merge-base) diff using: tpch File an issue against this benchmark runner |
|
🤖 Benchmark running (GKE) | trigger CPU Details (lscpu)Comparing worktree-io-uring-object-store (87539f6) to 7bfa3fb (merge-base) diff using: tpcds File an issue against this benchmark runner |
|
🤖 Benchmark running (GKE) | trigger CPU Details (lscpu)Comparing worktree-io-uring-object-store (87539f6) to 7bfa3fb (merge-base) diff using: clickbench_partitioned File an issue against this benchmark runner |
|
🤖 Benchmark completed (GKE) | trigger Instance: CPU Details (lscpu)Details
Resource Usagetpcds — base (merge-base)
tpcds — branch
File an issue against this benchmark runner |
|
🤖 Benchmark completed (GKE) | trigger Instance: CPU Details (lscpu)Details
Resource Usageclickbench_partitioned — base (merge-base)
clickbench_partitioned — branch
File an issue against this benchmark runner |
- Remove unnecessary `as u64` cast (meta.size is already u64 in 0.13.2) - Allow clippy::result_large_err on execute_read_ranges (object_store::Error is large by design) - Fix broken rustdoc links to LocalFileSystem (cfg-gated away when io-uring feature is enabled) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Dandandan
added a commit
to Dandandan/datafusion-benchmarking
that referenced
this pull request
Apr 16, 2026
DataFusion now includes io_uring-based local file I/O (apache/datafusion#21673), but the default container seccomp profile blocks io_uring_setup. Adds an init container that writes a seccomp profile (default allowlist + io_uring_setup/enter/register) to the node's kubelet seccomp dir. The runner container then references it via Localhost seccomp type. No DaemonSet or cluster-wide changes needed — it's self-contained in each benchmark Job.
Dandandan
added a commit
to Dandandan/datafusion-benchmarking
that referenced
this pull request
Apr 16, 2026
DataFusion now includes io_uring-based local file I/O (apache/datafusion#21673), but the default container seccomp profile blocks io_uring_setup. Adds a checked-in seccomp profile (services/seccomp/io-uring-allowed.json) that's the standard containerd default allowlist plus three syscalls: io_uring_setup, io_uring_enter, io_uring_register. Deployment: - A Pulumi DaemonSet (services/seccomp.ts) copies the profile from a ConfigMap to /var/lib/kubelet/seccomp/profiles/ on every node - The controller and benchmark-main workflow reference it via seccompProfile.type: Localhost No init containers needed — the DaemonSet keeps the profile present on all nodes continuously.
Dandandan
added a commit
to Dandandan/datafusion-benchmarking
that referenced
this pull request
Apr 16, 2026
DataFusion now includes io_uring-based local file I/O (apache/datafusion#21673), but the default container seccomp profile blocks io_uring_setup. Adds a checked-in seccomp profile (services/seccomp/io-uring-allowed.json) that's the standard containerd default allowlist plus three syscalls: io_uring_setup, io_uring_enter, io_uring_register. Deployment: - A Pulumi DaemonSet (services/seccomp.ts) copies the profile from a ConfigMap to /var/lib/kubelet/seccomp/profiles/ on every node - The controller and benchmark-main workflow reference it via seccompProfile.type: Localhost No init containers needed — the DaemonSet keeps the profile present on all nodes continuously.
adriangb
added a commit
to adriangb/datafusion-benchmarking
that referenced
this pull request
Apr 17, 2026
DataFusion now includes io_uring-based local file I/O (apache/datafusion#21673), but the default container seccomp profile blocks io_uring_setup. Adds a checked-in seccomp profile (services/seccomp/io-uring-allowed.json) that's the standard containerd default allowlist plus three syscalls: io_uring_setup, io_uring_enter, io_uring_register. Deployment: - A Pulumi DaemonSet (services/seccomp.ts) copies the profile from a ConfigMap to /var/lib/kubelet/seccomp/profiles/ on every node - The controller and benchmark-main workflow reference it via seccompProfile.type: Localhost No init containers needed — the DaemonSet keeps the profile present on all nodes continuously. Co-authored-by: Adrian Garcia Badaracco <1755071+adriangb@users.noreply.github.com>
Contributor
Author
|
run benchmarks |
1 similar comment
Contributor
Author
|
run benchmarks |
|
🤖 Benchmark running (GKE) | trigger CPU Details (lscpu)Comparing worktree-io-uring-object-store (5df85eb) to 7bfa3fb (merge-base) diff using: tpcds File an issue against this benchmark runner |
|
🤖 Benchmark running (GKE) | trigger CPU Details (lscpu)Comparing worktree-io-uring-object-store (5df85eb) to 7bfa3fb (merge-base) diff using: tpch File an issue against this benchmark runner |
|
🤖 Benchmark running (GKE) | trigger CPU Details (lscpu)Comparing worktree-io-uring-object-store (5df85eb) to 7bfa3fb (merge-base) diff using: clickbench_partitioned File an issue against this benchmark runner |
|
🤖 Benchmark completed (GKE) | trigger Instance: CPU Details (lscpu)Details
Resource Usagetpcds — base (merge-base)
tpcds — branch
File an issue against this benchmark runner |
|
🤖 Benchmark completed (GKE) | trigger Instance: CPU Details (lscpu)Details
Resource Usageclickbench_partitioned — base (merge-base)
clickbench_partitioned — branch
File an issue against this benchmark runner |
Dandandan
pushed a commit
to Dandandan/arrow-datafusion
that referenced
this pull request
Apr 17, 2026
Mirror the `datafusion-object-store-iouring` crate and registry wiring from apache#21673 into this branch so we can iterate on the io_uring path here.
Contributor
Author
|
run benchmarks |
|
🤖 Benchmark running (GKE) | trigger CPU Details (lscpu)Comparing worktree-io-uring-object-store (5df85eb) to 7bfa3fb (merge-base) diff using: tpch File an issue against this benchmark runner |
|
🤖 Benchmark running (GKE) | trigger CPU Details (lscpu)Comparing worktree-io-uring-object-store (5df85eb) to 7bfa3fb (merge-base) diff using: clickbench_partitioned File an issue against this benchmark runner |
|
🤖 Benchmark running (GKE) | trigger CPU Details (lscpu)Comparing worktree-io-uring-object-store (5df85eb) to 7bfa3fb (merge-base) diff using: tpcds File an issue against this benchmark runner |
Roll back the `uring::is_available()` probe and the `Option<UnboundedSender>` machinery added in 87539f6 so that the store is back to unconditionally using io_uring on Linux. Silently falling back to `LocalFileSystem` when io_uring_setup is blocked (EPERM, seccomp, etc.) hides real configuration problems. Follow-up commits address the correctness and performance gaps in the io_uring path itself.
- Path resolution: `self.root.join(location.as_ref())` joined the raw percent-encoded `object_store::Path` string, which diverges from the `LocalFileSystem` prefix/decoding logic on roots that aren't `/` and on any segment containing characters `object_store` encodes. Route resolution through `LocalFileSystem::path_to_filesystem` so puts and io_uring reads agree. - Partial reads: the drain loop previously `truncate(bytes_read)`'d on the first completion, silently returning short data. Track per-range fill progress and resubmit reads until the full range is covered or the kernel reports an error / unexpected EOF. - CQE bounds: `user_data` is cast to `usize` and used to index the buffer vector. Reject user_data values outside the buffer range instead of panicking or reading past the allocation. - CQE drain wait: if fewer than the expected number of CQEs are available after the initial `submit_and_wait`, wait for the remainder instead of busy-spinning on `completion()`. - `GetOptions` conditionals: `if_match`, `if_none_match`, `if_modified_since`, `if_unmodified_since`, and `version` were silently dropped. Send them to `LocalFileSystem::get_opts` with `head = true` so preconditions are enforced before we issue io_uring reads. - Panics: constructors now return `object_store::Result<Self>` instead of `expect`ing on invalid prefixes or thread-spawn failures; the SQ push retry returns an error on repeated failure rather than `expect`ing. - Misc: reject ranges with `start > end`, reject range lengths that exceed `usize`, and skip io_uring-backed unit tests cleanly when `io_uring_setup` is blocked by the sandbox.
The previous worker popped one `IoCommand` at a time, did
`submit_and_wait(n)` for its ranges, drained, and only then looked at
the next command. Concurrent async callers were effectively serialized
through that single `submit_and_wait`, negating io_uring's main win —
overlapping in-flight reads.
Rework the worker around a slab of `PendingRequest`s keyed by a
`u32` slot id. SQE `user_data` is `(slot << 32) | range_idx`, so every
completion finds its owning request and buffer directly. The loop:
1. Pull new commands until `MAX_IN_FLIGHT_REQUESTS` is reached (or
block on the channel when nothing is in flight).
2. Push queued SQEs into the ring until it is full.
3. `submit_and_wait(1)` when work is outstanding, then drain all
available CQEs. Partial reads re-queue a follow-up SQE.
Slots are only freed once `sqes_outstanding == 0`, so a late CQE can
never scribble into a reallocated slot's buffers.
Also:
* Switch the command channel to a bounded `mpsc::channel(1024)` so
bursts of requests apply backpressure rather than ballooning memory.
* Drop the redundant in-function chunking loop — the slab + backlog
already handles overflow.
* Add a multi-threaded integration test that fires 16 concurrent
`get_ranges` calls to assert the pipelining does not deadlock and
returns correct data.
Contributor
Author
|
run benchmarks |
|
🤖 Benchmark running (GKE) | trigger CPU Details (lscpu)Comparing worktree-io-uring-object-store (7fdb92b) to 7bfa3fb (merge-base) diff using: tpch File an issue against this benchmark runner |
|
🤖 Benchmark running (GKE) | trigger CPU Details (lscpu)Comparing worktree-io-uring-object-store (7fdb92b) to 7bfa3fb (merge-base) diff using: clickbench_partitioned File an issue against this benchmark runner |
|
🤖 Benchmark running (GKE) | trigger CPU Details (lscpu)Comparing worktree-io-uring-object-store (7fdb92b) to 7bfa3fb (merge-base) diff using: tpcds File an issue against this benchmark runner |
|
🤖 Benchmark completed (GKE) | trigger Instance: CPU Details (lscpu)Details
Resource Usagetpcds — base (merge-base)
tpcds — branch
File an issue against this benchmark runner |
|
🤖 Benchmark completed (GKE) | trigger Instance: CPU Details (lscpu)Details
Resource Usageclickbench_partitioned — base (merge-base)
clickbench_partitioned — branch
File an issue against this benchmark runner |
|
🤖 Benchmark completed (GKE) | trigger Instance: CPU Details (lscpu)Details
Resource Usagetpcds — base (merge-base)
tpcds — branch
File an issue against this benchmark runner |
|
🤖 Benchmark completed (GKE) | trigger Instance: CPU Details (lscpu)Details
Resource Usageclickbench_partitioned — base (merge-base)
clickbench_partitioned — branch
File an issue against this benchmark runner |
Switch the command channel to unbounded so senders never await, and trim hot-path overhead in the worker: - VecDeque backlog avoids the O(N) drain of the front. - Buffers grow via set_len as completions arrive, removing the zero-fill cost of vec![0u8; n] without reading uninit memory. - submit_and_wait(1) alone both flushes SQEs and waits, saving a syscall per loop iteration. - Drop the fixed in-flight cap; backpressure now comes from the ring and CQ drain. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Contributor
Author
|
run benchmarks |
|
🤖 Benchmark running (GKE) | trigger CPU Details (lscpu)Comparing worktree-io-uring-object-store (da7abc7) to 7bfa3fb (merge-base) diff using: clickbench_partitioned File an issue against this benchmark runner |
|
🤖 Benchmark running (GKE) | trigger CPU Details (lscpu)Comparing worktree-io-uring-object-store (da7abc7) to 7bfa3fb (merge-base) diff using: tpcds File an issue against this benchmark runner |
|
🤖 Benchmark running (GKE) | trigger CPU Details (lscpu)Comparing worktree-io-uring-object-store (da7abc7) to 7bfa3fb (merge-base) diff using: tpch File an issue against this benchmark runner |
|
🤖 Benchmark completed (GKE) | trigger Instance: CPU Details (lscpu)Details
Resource Usagetpcds — base (merge-base)
tpcds — branch
File an issue against this benchmark runner |
|
🤖 Benchmark completed (GKE) | trigger Instance: CPU Details (lscpu)Details
Resource Usageclickbench_partitioned — base (merge-base)
clickbench_partitioned — branch
File an issue against this benchmark runner |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Which issue does this PR close?
Rationale for this change
Local file reads in DataFusion use
object_store::local::LocalFileSystem, which issues onepread()per byte range. For Parquet column chunk reads this means many individual syscalls. Linux's io_uring allows batching these into a singleio_uring_enter().What changes are included in this PR?
New crate
datafusion-object-store-iouringprovidingIoUringObjectStore:get_ranges()submits all byte ranges as SQEs in oneio_uring_enter()LocalFileSystemwhen unavailable (EPERM in Docker/seccomp, non-Linux platforms)io-uringondatafusion-executionanddatafusion, enabled by defaultArchitecture:
To actually enable io_uring in benchmark environments, a companion PR adds a custom seccomp profile: adriangb/datafusion-benchmarking#4.
Are these changes tested?
6 unit tests covering put/get round-trip, single and multi-range reads, head, list, and empty-range edge cases. The io_uring code path requires a Linux host with io_uring support; on macOS CI the fallback path is exercised.
Are there any user-facing changes?
New optional crate and default feature flag. No behavioral changes when io_uring is unavailable — transparent fallback to
LocalFileSystem.