Run partial Flat random access with batched io_uring

.github/workflows/pr-bench-runner.yml

@@ -98,6 +98,11 @@ jobs:
           RUST_BACKTRACE: full
           VORTEX_EXPERIMENTAL_PATCHED_ARRAY: "1"
           FLAT_LAYOUT_INLINE_ARRAY_NODE: "1"
+          VORTEX_IO_URING: "1"
+          VORTEX_IO_URING_RINGS: "4"
+          VORTEX_IO_URING_QUEUE_DEPTH: "128"
+          VORTEX_IO_URING_MIN_READ_SIZE: "0"
+          VORTEX_IO_URING_MAX_IN_FLIGHT: "512"
         run: |
           python3 scripts/random-access-split.py
 

Cargo.toml

@@ -169,6 +169,7 @@ geoarrow = "0.8.0"
 geoarrow-cast = "0.8.0"
 get_dir = "0.5.0"
 glob = "0.3.2"
+io-uring = "0.7.13"
 goldenfile = "1"
 half = { version = "2.7.1", features = ["std", "num-traits"] }
 hashbrown = "0.17.1"

benchmark-results/pr-9416-random-access/HANDOVER.md

@@ -0,0 +1,88 @@
+# PR #9416 random-access handover
+
+## Scope
+
+- PR: https://github.com/vortex-data/vortex/pull/9416
+- Branch: `ji/partial-flat-random-access-batched-ring`
+- Worktree: `/mnt/vortex-ssd/worktrees/random-access-pr9416-ring`
+- Workload: feature-vectors / uniform, pinned to cores 0-7.
+- The decoded ALPRD patch cache was removed. Do not reintroduce a data cache for the I/O comparison.
+- The `action/bench-random-access` label was applied after removing the cache.
+
+## Retained change
+
+The ALPRD partial-read path no longer calls `clear_stats` on the newly constructed partial array.
+Those `BitPacked`, `Patches`, `ALPRD`, and `FixedSizeList` arrays begin with empty statistics. The
+fixed-width path still clears statistics inherited from its serialized array tree.
+
+Three five-second cached runs on 8 cores:
+
+| Variant | Vortex runs | Median | Lance median |
+|---|---|---:|---:|
+| No-cache baseline | 1.612 / 1.621 / 1.680 ms | 1.621 ms | 1.059 ms |
+| Skip redundant ALPRD stats clear | 1.594 / 1.602 / 1.607 ms | 1.601 ms | 1.058 ms |
+
+This is a 1.2% Vortex improvement and changes neither I/O nor serialization. A final cold rerun is
+still desirable, although the change occurs entirely after I/O.
+
+## Segment and patch measurements
+
+The complete 105-row measurement is in `feature-vectors-uniform-segments.csv`.
+
+- Uniform selects 105 physical Flat segments.
+- Every segment has 256 vectors x 1,024 values = 262,144 values.
+- Left buffer: 98,304 bytes per segment; right buffer: 753,664 bytes per segment.
+- Patches per segment: min 953, mean 1,060.70, max 1,793.
+- Patch density: 0.404627%.
+- Patch bytes per segment: min 5,718, mean 6,364.23, max 10,758.
+- Totals: 27,525,120 resident values, 111,374 patches, 668,244 patch bytes.
+- The query selects 107,520 values and should intersect only about 435 patches, but the reader
+  reconstructs all 111,374 patches before slicing.
+- Main data is already row-sliced: 384 left bytes + 2,944 right bytes per selected vector, or
+  349,440 bytes total. Patch buffers remain unsliced and add 668,244 bytes.
+
+Conclusion: 256-vector Flat segments are not the main problem because left/right buffers support
+partial row reads. Patch read/reconstruction granularity is the leak.
+
+## Rejected experiments
+
+1. `OnceLock<Patches>` cache: hot 1.491 ms, but skipped patch reads and decode on repeated takes.
+   Removed as an invalid apples-to-apples I/O comparison.
+2. Search bitpacked patch indices without bulk decode: median 1.640 ms versus 1.621 ms baseline.
+   The scalar probes were slower; removed.
+3. Merge patch-index/value reads: median 1.741 ms. `FileSegmentSource` already coalesces nearby
+   reads, so this added slicing work; removed, including the temporary buffer API.
+4. Latency-based blocking/io_uring routing: about 1.28 ms hot but about 26.7 ms cold because a small
+   cold metadata read falsely classified the file as hot; removed.
+
+## Profile and next work
+
+A 10-second no-cache Samply profile kept all eight Tokio workers busy. Dominant resolved self
+frames were AArch64 atomics and mimalloc allocation/free, pointing to task/array/future construction
+overhead rather than worker serialization. The local profile is
+`/tmp/no-cache-feature-uniform.profile.json.gz` and is not portable with this branch.
+
+Next, count allocations and short-lived objects inside `resolve_alprd_pages` and final
+canonicalization. Avoid two-stage patch I/O unless separately proven hot and cold: it can reduce
+patch-value bytes but adds an I/O round. If a future format change is allowed, serialize ALPRD patch
+chunk offsets. The 1,024-value patch chunk exactly matches one feature vector, enabling direct
+lookup of the roughly four relevant patches.
+
+## Benchmark environment
+
+```text
+taskset -c 0-7
+TOKIO_WORKER_THREADS=8
+RAYON_NUM_THREADS=8
+LANCE_IO_THREADS=8
+VORTEX_EXPERIMENTAL_PATCHED_ARRAY=1
+FLAT_LAYOUT_INLINE_ARRAY_NODE=1
+VORTEX_IO_URING=1
+VORTEX_IO_URING_RINGS=4
+VORTEX_IO_URING_QUEUE_DEPTH=128
+VORTEX_IO_URING_MIN_READ_SIZE=0
+VORTEX_IO_URING_MAX_IN_FLIGHT=512
+```
+
+Validation completed: nightly formatting, `git diff --check`, `cargo check -p vortex-layout`,
+release benchmark builds, feature-vectors/uniform smoke runs, and repeated hot comparisons.

benchmark-results/pr-9416-random-access/feature-vectors-uniform-segments.csv

@@ -0,0 +1,106 @@
+segment_id,rows,values,patches,patch_bytes,left_bytes,right_bytes
+26,256,262144,1031,6186,98304,753664
+62,256,262144,1018,6108,98304,753664
+81,256,262144,994,5964,98304,753664
+98,256,262144,993,5958,98304,753664
+99,256,262144,1011,6066,98304,753664
+119,256,262144,1007,6042,98304,753664
+124,256,262144,1017,6102,98304,753664
+220,256,262144,1022,6132,98304,753664
+231,256,262144,1018,6108,98304,753664
+232,256,262144,1049,6294,98304,753664
+337,256,262144,1041,6246,98304,753664
+439,256,262144,1067,6402,98304,753664
+475,256,262144,985,5910,98304,753664
+479,256,262144,1000,6000,98304,753664
+482,256,262144,1038,6228,98304,753664
+495,256,262144,1034,6204,98304,753664
+502,256,262144,1027,6162,98304,753664
+503,256,262144,1039,6234,98304,753664
+595,256,262144,1022,6132,98304,753664
+627,256,262144,1541,9246,98304,753664
+686,256,262144,987,5922,98304,753664
+704,256,262144,1016,6096,98304,753664
+712,256,262144,988,5928,98304,753664
+744,256,262144,1051,6306,98304,753664
+785,256,262144,980,5880,98304,753664
+808,256,262144,1054,6324,98304,753664
+815,256,262144,1008,6048,98304,753664
+821,256,262144,1014,6084,98304,753664
+840,256,262144,1089,6534,98304,753664
+849,256,262144,1015,6090,98304,753664
+865,256,262144,1030,6180,98304,753664
+908,256,262144,995,5970,98304,753664
+1063,256,262144,999,5994,98304,753664
+1081,256,262144,1775,10650,98304,753664
+1107,256,262144,1067,6402,98304,753664
+1131,256,262144,1027,6162,98304,753664
+1143,256,262144,1573,9438,98304,753664
+1148,256,262144,1051,6306,98304,753664
+1187,256,262144,1068,6408,98304,753664
+1227,256,262144,1041,6246,98304,753664
+1329,256,262144,993,5958,98304,753664
+1354,256,262144,973,5838,98304,753664
+1356,256,262144,1003,6018,98304,753664
+1374,256,262144,1045,6270,98304,753664
+1406,256,262144,963,5778,98304,753664
+1436,256,262144,1056,6336,98304,753664
+1460,256,262144,1000,6000,98304,753664
+1573,256,262144,953,5718,98304,753664
+1599,256,262144,996,5976,98304,753664
+1632,256,262144,1008,6048,98304,753664
+1667,256,262144,1033,6198,98304,753664
+1763,256,262144,1067,6402,98304,753664
+1778,256,262144,1014,6084,98304,753664
+1786,256,262144,1002,6012,98304,753664
+1816,256,262144,1017,6102,98304,753664
+1834,256,262144,1048,6288,98304,753664
+1869,256,262144,1042,6252,98304,753664
+1910,256,262144,1039,6234,98304,753664
+1953,256,262144,1011,6066,98304,753664
+1993,256,262144,1061,6366,98304,753664
+2010,256,262144,1078,6468,98304,753664
+2071,256,262144,1011,6066,98304,753664
+2091,256,262144,1047,6282,98304,753664
+2096,256,262144,1022,6132,98304,753664
+2146,256,262144,993,5958,98304,753664
+2176,256,262144,1030,6180,98304,753664
+2190,256,262144,1058,6348,98304,753664
+2236,256,262144,980,5880,98304,753664
+2238,256,262144,1064,6384,98304,753664
+2270,256,262144,1053,6318,98304,753664
+2282,256,262144,997,5982,98304,753664
+2389,256,262144,1064,6384,98304,753664
+2447,256,262144,1000,6000,98304,753664
+2484,256,262144,1046,6276,98304,753664
+2501,256,262144,1050,6300,98304,753664
+2567,256,262144,999,5994,98304,753664
+2569,256,262144,1051,6306,98304,753664
+2585,256,262144,1018,6108,98304,753664
+2639,256,262144,1507,9042,98304,753664
+2659,256,262144,1013,6078,98304,753664
+2694,256,262144,1533,9198,98304,753664
+2763,256,262144,1078,6468,98304,753664
+2842,256,262144,1067,6402,98304,753664
+2981,256,262144,1037,6222,98304,753664
+3001,256,262144,982,5892,98304,753664
+3007,256,262144,1018,6108,98304,753664
+3036,256,262144,990,5940,98304,753664
+3057,256,262144,973,5838,98304,753664
+3067,256,262144,962,5772,98304,753664
+3068,256,262144,1016,6096,98304,753664
+3219,256,262144,994,5964,98304,753664
+3298,256,262144,1034,6204,98304,753664
+3321,256,262144,983,5898,98304,753664
+3384,256,262144,1029,6174,98304,753664
+3410,256,262144,1012,6072,98304,753664
+3422,256,262144,1028,6168,98304,753664
+3578,256,262144,1032,6192,98304,753664
+3622,256,262144,1101,6606,98304,753664
+3637,256,262144,1036,6216,98304,753664
+3639,256,262144,975,5850,98304,753664
+3682,256,262144,1793,10758,98304,753664
+3760,256,262144,996,5976,98304,753664
+3849,256,262144,1480,8880,98304,753664
+3851,256,262144,999,5994,98304,753664
+3865,256,262144,1039,6234,98304,753664

benchmarks/datafusion-bench/src/main.rs

@@ -289,10 +289,14 @@ async fn register_v2_tables<B: Benchmark + ?Sized>(
             .runtime_env()
             .object_store(table_url.object_store())?;
 
-        let fs: FileSystemRef = Arc::new(ObjectStoreFileSystem::new(
-            Arc::clone(&store),
-            SESSION.handle(),
-        ));
+        let fs: FileSystemRef = if benchmark_base.scheme() == "file" {
+            Arc::new(ObjectStoreFileSystem::local(SESSION.handle()))
+        } else {
+            Arc::new(ObjectStoreFileSystem::new(
+                Arc::clone(&store),
+                SESSION.handle(),
+            ))
+        };
         let base_prefix = benchmark_base.path().trim_start_matches('/').to_string();
         let fs = fs.with_prefix(base_prefix);
 

encodings/alp/src/alp_rd/array.rs

@@ -71,6 +71,33 @@ pub struct ALPRDMetadata {
     patches: Option<PatchesMetadata>,
 }
 
+impl ALPRDMetadata {
+    pub fn right_bit_width(&self) -> VortexResult<u8> {
+        u8::try_from(self.right_bit_width).map_err(|_| {
+            vortex_err!(
+                "right bit width {} does not fit in u8",
+                self.right_bit_width
+            )
+        })
+    }
+
+    pub fn left_parts_dictionary(&self) -> VortexResult<Buffer<u16>> {
+        self.dict
+            .get(..usize::try_from(self.dict_len)?)
+            .ok_or_else(|| vortex_err!("ALPRD dictionary length is out of bounds"))?
+            .iter()
+            .map(|&value| {
+                u16::try_from(value)
+                    .map_err(|_| vortex_err!("ALPRD dictionary value {value} does not fit in u16"))
+            })
+            .collect()
+    }
+
+    pub fn patches(&self) -> Option<&PatchesMetadata> {
+        self.patches.as_ref()
+    }
+}
+
 impl ArrayHash for ALPRDData {
     fn array_hash<H: Hasher>(&self, state: &mut H, accuracy: EqMode) {
         self.left_parts_dictionary.array_hash(state, accuracy);

encodings/fastlanes/src/bitpacking/mod.rs

@@ -19,6 +19,7 @@ mod vtable;
 pub(crate) use plugin::BitPackedPatchedPlugin;
 pub use vtable::BitPacked;
 pub use vtable::BitPackedArray;
+pub use vtable::BitPackedMetadata;
 
 pub(crate) fn initialize(session: &vortex_session::VortexSession) {
     vtable::initialize(session);

encodings/fastlanes/src/bitpacking/vtable/mod.rs

@@ -72,6 +72,22 @@ pub struct BitPackedMetadata {
     pub(crate) patches: Option<PatchesMetadata>,
 }
 
+impl BitPackedMetadata {
+    pub fn bit_width(&self) -> VortexResult<u8> {
+        u8::try_from(self.bit_width)
+            .map_err(|_| vortex_err!("bit width {} does not fit in u8", self.bit_width))
+    }
+
+    pub fn offset(&self) -> VortexResult<u16> {
+        u16::try_from(self.offset)
+            .map_err(|_| vortex_err!("bit-packed offset {} does not fit in u16", self.offset))
+    }
+
+    pub fn patches(&self) -> Option<&PatchesMetadata> {
+        self.patches.as_ref()
+    }
+}
+
 impl ArrayHash for BitPackedData {
     fn array_hash<H: Hasher>(&self, state: &mut H, accuracy: EqMode) {
         self.offset.hash(state);

vortex-array/src/arrays/list/mod.rs

@@ -14,6 +14,7 @@ pub(crate) mod compute;
 
 mod vtable;
 pub use vtable::List;
+pub use vtable::ListMetadata;
 
 pub(crate) fn initialize(session: &vortex_session::VortexSession) {
     compute::initialize(session);

vortex-array/src/arrays/list/vtable/mod.rs

@@ -52,6 +52,12 @@ pub struct ListMetadata {
     offset_ptype: i32,
 }
 
+impl ListMetadata {
+    pub fn elements_len(&self) -> u64 {
+        self.elements_len
+    }
+}
+
 impl ArrayHash for ListData {
     fn array_hash<H: Hasher>(&self, _state: &mut H, _accuracy: EqMode) {}
 }