feat: support Apache Paimon non-direct-file reads by yanbinyang · Pull Request #616 · milvus-io/milvus-storage

yanbinyang · 2026-08-06T04:16:35Z

Part of #593

Design + tracking: milvus-io/milvus#45881

Related tracking: milvus-io/milvus#50632

Consumed by: milvus-io/milvus#51897

What changed

Add native Apache Paimon DataSplit reads as a follow-up to #602.

Persist and validate versioned native DataSplit descriptors using paimon-rust 0.3.0.
Stream projected Arrow batches without materializing a complete split.
Stream each DataSplit read operation without materializing the complete split, while keeping the forward-only adapter inside the Paimon reader.
Add the data-split scan mode and automatic fallback when direct-file reads cannot preserve Paimon semantics.
Add Loon fixtures for append-only, deletion-vector, and merge-on-read tables.

Routing

auto keeps safe raw-convertible Parquet and Vortex splits on the direct-file path.
auto falls back to DataSplit for merge-on-read, row ranges, partition columns, schema or data evolution, sidecars, and unsupported direct-file formats.
direct-file fails explicitly when a split cannot be read safely.
data-split forces native Paimon execution.

The direct-file behavior introduced by #602 remains unchanged.

Compared with the earlier combined #594 implementation, this PR is limited to DataSplit execution, uses the released native split codec, avoids whole-split materialization, coalesces normal eager and lazy loads onto a forward stream, and isolates take, range, and backward reads from the shared load cursor.

Test plan

Rust unit tests
Paimon C++ integration and error-classification tests
Loon append-only, deletion-vector, and merge-on-read scenarios
Milvus external-collection Go unit tests
Milvus standalone client/v3 E2E with auto, direct-file, DataSplit, deletion-vector, merge-on-read, and pinned snapshots

Signed-off-by: YangYanbin warlock.yyb@alibaba-inc.com

sre-ci-robot · 2026-08-06T04:16:40Z

[APPROVALNOTIFIER] This PR is NOT APPROVED

This pull-request has been approved by: yanbinyang
To complete the pull request process, please assign sunby after the PR has been reviewed.
You can assign the PR to them by writing /assign @sunby in a comment when ready.

The full list of commands accepted by this bot can be found here.

Details

Needs approval from an approver in each of these files:

OWNERS

Approvers can indicate their approval by writing /approve in a comment
Approvers can cancel approval by writing /approve cancel in a comment

codecov · 2026-08-06T04:38:38Z

Codecov Report

❌ Patch coverage is 88.07692% with 31 lines in your changes missing coverage. Please review.
✅ Project coverage is 76.58%. Comparing base (1fd5edf) to head (31c3bf1).
⚠️ Report is 4 commits behind head on main.

Files with missing lines	Patch %	Lines
cpp/src/format/paimon/paimon_format_reader.cpp	88.53%	29 Missing ⚠️
cpp/src/format/bridge/rust/include/paimon_bridge.h	0.00%	1 Missing ⚠️
cpp/src/format/paimon/paimon_format.cpp	83.33%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #616      +/-   ##
==========================================
+ Coverage   76.19%   76.58%   +0.39%     
==========================================
  Files         173      173              
  Lines       17388    17747     +359     
  Branches     2618     2699      +81     
==========================================
+ Hits        13248    13591     +343     
- Misses       4140     4156      +16

Flag	Coverage Δ
cpp	`79.15% <88.07%> (+0.36%)`	⬆️
python	`44.45% <ø> (ø)`

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

Accept Paimon table specifications and propagate snapshot, scan-mode, and object-storage options through external collection planning. Route raw-compatible files through direct-file readers and preserve native DataSplits for scans that require Paimon execution. Keep DataSplits atomic, group them by bucket, and pack them into Milvus segments with the existing row target and a configurable split-count limit. Preserve storage-owned descriptors in StorageV3 manifests and use decoded-memory estimates for load planning. Add C++, Go, and client coverage for routing, manifest round trips, segment packing, deletion vectors, merge-on-read, and snapshot reads. Storage direct-file support: milvus-io/milvus-storage#602. Depends on: milvus-io/milvus-storage#616. Design + tracking: milvus-io#45881. Related tracking: milvus-io#50632. --------- Signed-off-by: YangYanbin <warlock.yyb@alibaba-inc.com>

jiaqizho · 2026-08-07T02:08:57Z

-  auto all_tasks = ChunkTask::Build(unique_chunk_indices, get_chunk_info);
-  SplitAsyncTasks(all_tasks, std::max<size_t>(parallelism, 1), ChunkTask::SplitTraits{get_chunk_info},
-                  GetAsyncTaskSplitStrategy(properties_));
+  auto natural_tasks = ChunkTask::Build(unique_chunk_indices, get_chunk_info);


Currently, Paimon does not implement the native async interface. It uses the generic fallback, which executes the blocking read and wraps the result in an async task/future. Given that, I think the sequential-read constraint should remain within Paimon’s own get_chunk semantics instead of changing the common ChunkTask splitting logic.

Also, for the non-direct-read path, the current approach does not fully solve the duplicated-read problem. I would prefer to keep this PR focused on functional correctness first, and address the performance issue later with a more complete design.

OK. I’ll simplify this PR to focus on functional correctness, remove the cross-call cursor reuse and related test hooks, and leave broader DataSplit read optimization for a separate change.

jiaqizho · 2026-08-07T02:11:16Z

+  // Forward-only readers must receive requested chunks for one file through a
+  // shared ascending cursor. This avoids reopening and replaying their source
+  // stream for every logical range.
+  [[nodiscard]] virtual bool requires_sequential_chunk_reads() const { return false; }


I don’t think requires_sequential_chunk_reads is a good API for adapting a forward-only streaming reader to a random-access chunk interface.

Removed in 4283be8d. The common reader and ChunkTask scheduling are unchanged; Paimon now creates a private forward cursor for each read operation.

jiaqizho · 2026-08-07T02:22:04Z

change commit title to feat: support Apache Paimon non-direct-file reads ?

jiaqizho · 2026-08-07T02:37:06Z

-      if (metadata.value("read_path", std::string{}) != "direct-file") {
-        return arrow::Status::Invalid("Paimon planner returned a non-direct metadata descriptor");
-      }
+      read_path = metadata.value("read_path", std::string{});


constexpr "read_path"/direct-file/data-split

and check others

Updated in 4283be8d. read_path, direct-file, and data-split now use shared constants from paimon_common.h.

yhmo · 2026-08-07T04:49:17Z

+    schema: SchemaRef,
+}
+
+fn classify_stream_error(error: paimon::Error) -> ArrowError {


cpp/src/format/bridge/rust/src/paimon_bridgeimpl.rs line:999
Medium ---- This stream-specific mapping turns every non-temporary IoUnexpected (including NotFound and PermissionDenied) into InvalidArgumentError, while temporary failures become a generic Arrow I/O error and lose the StorageTransientThrottling / StorageTransientService details used by the rest of the bridge. A failure after stream open therefore has different retry and refresh semantics from planning/open. Please preserve the same not-found, transient, and plain-I/O classification across the Arrow stream boundary.

Fixed in c70fa849. Stream errors now preserve the existing Paimon markers through Arrow FFI, so not-found, transient, invalid, not-implemented, and plain I/O errors keep the same handling as planning and open failures. The FFI and C++ status mappings are covered by tests.

yhmo · 2026-08-07T04:49:18Z

+    // Keep the handle: every reader created from this cached metadata shares
+    // it, so schema and snapshot resolution happen once per descriptor.
+    metadata->payload.split_reader_handle = std::move(reader);
+    ARROW_ASSIGN_OR_RAISE(metadata->row_group_infos, MakeLogicalRowGroups(parsed.record_count, logical_chunk_rows));


cpp/src/format/paimon/paimon_format_reader.cpp line:557
Medium ---- Planning filters out zero-row DataSplits, but the persisted descriptor reader accepts record_count == 0 and MakeLogicalRowGroups publishes no chunks. Because no stream is then consumed, the EOF validation never compares the declared count with the actual split, so corrupted metadata can silently hide a non-empty split. Please reject zero for data-split descriptors, or otherwise force explicit count validation before exposing the metadata.

Fixed in e8758f0e. Milvus planning already omits zero-row splits, so a persisted zero-row descriptor violates that contract. This commit rejects it during metadata parsing, preventing an empty chunk list from bypassing stream count validation. The new test covers a non-empty split whose declared count is corrupted to zero; the existing fully-deleted-table test verifies legitimate empty results are omitted during planning.

yhmo · 2026-08-07T04:49:19Z

    std::cerr << std::endl;
-    std::cerr << "Creates a demo table with schema (id int64, name string,"
-              << " value float64)." << std::endl;
+    std::cerr << "Scalar schema: id int64, name string, value float64." << std::endl;


cpp/tools/loon.cpp line:167
Low ---- This help now applies to both Iceberg and Paimon, but the Paimon scalar fixture declares id as Int32 in both fixture_schema and scalar_fixture_batch, not int64. A user or downstream E2E that follows this text can construct the wrong external schema. Please distinguish the two table types here or document Paimon's actual id type.

Fixed in 6f3a9a5f. The Paimon scalar fixture now uses BIGINT / Arrow Int64, so the shared Loon schema text is accurate. The dependent C++ and Loon tests were updated

Persist native Paimon DataSplit descriptors for scans that cannot be served safely through direct files, including merge-on-read, row-range, partitioned, and evolved tables. Stream projected Arrow batches without materializing a complete split. Keep forward-only split handling inside the Paimon reader so common sync and async scheduling remain unchanged; each chunk batch, take, and range operation owns its stream. Validate DataSplit bindings and declared row counts, preserve existing direct-file behavior, and cover automatic and forced routing, cross-batch reads, sparse take, and malformed descriptors. Signed-off-by: YangYanbin <warlock.yyb@alibaba-inc.com>

Generate append-only, deletion-vector, and merge-on-read Paimon tables through the existing demo-table command. Return committed snapshot history and accept storage options so external-collection E2E can use local or object-storage fixtures. Use int64 IDs for the scalar fixture to match the existing Iceberg and Lance test schemas. Also verify DataSplit deletion-vector application and merge-on-read values in the C++ integration suite. Signed-off-by: YangYanbin <warlock.yyb@alibaba-inc.com>

Carry the existing bridge error markers through Arrow C streams and translate them after import. Stream-time not-found and transient failures now retain the same retry semantics as planning and open failures, while plain storage failures remain I/O errors. Signed-off-by: YangYanbin <warlock.yyb@alibaba-inc.com>

Reject persisted data-split descriptors with a zero row count because planning never emits them and they otherwise bypass stream count validation. Signed-off-by: YangYanbin <warlock.yyb@alibaba-inc.com>

yanbinyang · 2026-08-07T16:51:07Z

change commit title to feat: support Apache Paimon non-direct-file reads ?

Updated the first commit title to feat: support Apache Paimon non-direct-file reads.

jiaqizho · 2026-08-10T08:09:09Z

+
+        unsafe fn open_stream(
+            self: &BlockingPaimonDataSplitReader,
+            projected_columns: &CxxVector<CxxString>,


Vec<String> ?

should convert vector to rust type in cpp.

open_stream consumes the projection synchronously and does not retain it, so I used a borrowed CxxVector to avoid copying the C++ strings into an owned Rust vector. Would you still prefer converting it on the C++ side to keep the Rust interface as Vec? I can change it if so.

for projection field copy, I do think that is fine.

Understood. I’ll switch the projection argument to an owned Vec and convert it on the C++ side.

Updated in 19d30d4.

jiaqizho · 2026-08-10T08:20:22Z

@@ -34,7 +37,38 @@
 namespace milvus_storage::paimon {
 namespace {


namespace { No need empty namespace. This is a bad habit of GPT.

The anonymous namespace contains helpers used only in this translation unit and keeps their symbols file-local, consistent with the Iceberg and Vortex readers in this repository. Would you prefer using static helpers here instead? I can adjust it if that is the intended style.

ok, just let me clean up empty namespaces across the entire storage project in the future.

jiaqizho · 2026-08-10T08:26:43Z

+            out_schema_ptr: *mut u8,
+        ) -> Result<()>;
+
+        unsafe fn open_stream(


can open_stream pass with footer or some meta?

Then different projection won't read footer in multi-times.

notice that: The external table will projects each column individually.

The current paimon-rust 0.3 API does not expose reusable file metadata across projected readers. I see two options:

Keep this PR on the released 0.3 API and accept repeated footer reads for now, then optimize it separately.

Add a scan-scoped read context with internal footer metadata caching to paimon-rust, and share it from BlockingPaimonDataSplitReader across projected streams. This is closer to paimon-cpp’s design, but requires upstream API work and likely waiting for paimon-rust 0.4.0, no earlier than the end of this month.

Would you prefer keeping this PR small, or waiting for the upstream API?

okay, So let's just leave a FIXME(name) then.

Updated in 19d30d4.

jiaqizho · 2026-08-10T08:31:29Z

+/// same cached metadata.
+class BlockingPaimonDataSplitReader {
+  public:
+  static arrow::Result<std::unique_ptr<BlockingPaimonDataSplitReader>> Open(const std::string& metadata_json,


direct return std::shared_ptr? Then u don't need warp it as shared

Updated in 400e0557.

jiaqizho · 2026-08-10T08:39:12Z

+    for (auto index : indices) {
+      const auto& group = metadata_->row_group_infos[index];
+      ARROW_ASSIGN_OR_RAISE(auto batches, cursor->ReadRange(group.start_offset, group.end_offset));
+      ARROW_ASSIGN_OR_RAISE(auto batch, CombineBatches(batches, output_schema_));


It will make copy.

In fact, the "logical chunk rows" value is merely an indicator rather than a hard requirement. You can observe the approach in vortex: it actually returns a native split to inform the caller how to partition the data so that no resulting RecordBatch exceeds the limit, thereby avoiding data copying.

We can allow chunks to be smaller, rather than copying them.

I checked Paimon Java and paimon-cpp. Neither exposes merged DataSplit output boundaries during planning. Both use a pull-based reader (readBatch / NextBatch) and allow each returned batch to have its actual size.
For DataSplit/MOR, output boundaries are determined during merge and deletion processing, so exposing physical file boundaries from paimon-rust would not describe the logical output. Removing CombineBatches cleanly would require a variable-sized streaming contract in milvus-storage, which is broader than this PR.

liliu-z · 2026-08-10T13:37:14Z

+  if (is_data_split()) {
+    ARROW_ASSIGN_OR_RAISE(auto cursor, make_data_split_cursor());
+    ARROW_ASSIGN_OR_RAISE(auto selected_batches, cursor->TakeRows(indices));
+    return arrow::Table::FromRecordBatches(output_schema_, selected_batches);


On the data-split path output_schema_ (which is simply read_schema whenever the caller supplies one, see ProjectSchema at :164) is bound to batches that carry Paimon's own stream schema, and nothing reconciles the two. CombineBatches returns the source batches untouched for any non-empty input, so get_chunk/get_chunks emit stream-schema batches and DataSplitRangeReader::schema() (:427) advertises a schema its ReadNext (:436) does not produce, while take() passes output_schema_ to arrow::Table::FromRecordBatches, which rejects any batch whose schema differs in type or nullability. A caller whose read_schema differs from Paimon's Arrow schema only in nullability therefore gets Invalid: Schema at index 0 was different from take() and from arrow::Table::FromRecordBatchReader on the range reader, even though the Paimon direct-file path and the other format readers accept the same input. Pick one rule for this path — project the stream onto output_schema_, or derive output_schema_ from the stream — and apply it to all three entry points. (Note: the original claim that take()/read_with_range() honor read_schema is not accurate; read_with_range uses the same CombineBatches, and take() only validates.)

Fixed in 6343a578. DataSplit batches are validated and rebound to output_schema_ for chunk, take, and range reads. Incompatible fields and nulls for non-nullable fields return Invalid.

liliu-z · 2026-08-10T13:37:15Z

+  auto column_groups = std::make_shared<api::ColumnGroups>();
+  column_groups->push_back(std::make_shared<api::ColumnGroup>(
+      api::ColumnGroup{.columns = {"id"}, .format = LOON_FORMAT_PAIMON_TABLE, .files = {files.front()}}));
+  auto schema = arrow::schema({arrow::field("id", arrow::int64())});


No new data-split test exercises take() or read_with_range() with a non-null read_schema: DataSplitTakeCompactsSparseBatches (:668) and DataSplitSupportsRangeAndCloneReads (:607) both build the reader with FormatReader::create(nullptr, ...), so output_schema_ is derived from Paimon's own stream schema and the schema binding cannot fail. The two tests that do pass a schema use arrow::schema({arrow::field("id", arrow::int64())}), which matches the fixture's nullable BIGINT exactly — and one of them only asserts total_rows() while the other goes through the chunk path. Add take/range cases with a read_schema that differs from Paimon's schema in nullability or type so the mismatch behaviour is pinned down rather than untested.

Covered in 6343a578. The test uses a non-null read_schema with different nullability for take and range reads, and verifies that an incompatible type is rejected.

liliu-z · 2026-08-10T13:37:17Z

+      }
+      previous = index;
+    }
+    ARROW_ASSIGN_OR_RAISE(auto cursor, make_data_split_cursor());


Every public read on the data-split path opens a brand-new Paimon merge stream and forward-skips to its start, and the reader keeps no cursor between calls. get_chunk(i) delegates to get_chunks({i}), which builds a fresh cursor, so iterating a split chunk-by-chunk decodes O(N^2) rows: for a 10M-row split at the default 8192-row chunk size that is ~1220 streams and ~6 billion rows decoded instead of 10 million. This is the primary Milvus access pattern — loon_get_chunk (cpp/src/ffi/reader_c.cpp:111) is per-chunk, and api::Reader itself falls back to single-chunk reads whenever a predicate is set (cpp/src/reader.cpp:449-457). Hoisting the cursor to a reader member and reusing it while start >= position_ (falling back to a fresh stream only for backward seeks) would keep the forward-only design and make sequential iteration linear; get_chunks already proves one cursor can serve many chunks.

Fixed in 81e4d6f. Sequential get_chunk calls reuse a reader-local forward cursor and reopen only for backward reads or after an error. Batch and range reads keep independent streams. The test covers both forward reuse and backward reopening.

A clarification to my earlier reply: the common reader and ChunkTask changes remain removed. Removed in 4283be8. After the later O(N²) issue was identified, 81e4d6f restores only a Paimon reader-local forward cursor for sequential get_chunk calls. It reopens for backward reads or errors, while batch and range reads use independent streams, so no common scheduling capability was reintroduced.

Parse demo-table row and dimension values without throwing on malformed input, and return the existing command error path instead. Signed-off-by: YangYanbin <warlock.yyb@alibaba-inc.com>

Recover not-found and transient storage classifications when the Paimon Parquet adapter stringifies FileRead errors, while leaving other permanent storage failures as I/O errors. Cover the Arrow stream path with Rust classification cases and a C++ mid-stream not-found test. Signed-off-by: YangYanbin <warlock.yyb@alibaba-inc.com>

Signed-off-by: YangYanbin <warlock.yyb@alibaba-inc.com>

Rebind projected batches to the requested Arrow schema when column names and types match, and reject incompatible schemas before they reach take or range consumers. Signed-off-by: YangYanbin <warlock.yyb@alibaba-inc.com>

Keep a reader-local forward cursor for get_chunk, and reopen it only for backward reads or failures. Batch and range operations continue to use independent streams. Signed-off-by: YangYanbin <warlock.yyb@alibaba-inc.com>

Expose the Paimon RecordBatchReader wrapper through the same internal test seam used by Vortex, and cover not-found and throttling status translation with synthetic readers. Delete data files before opening the integration stream so the test validates the final ENOENT contract without depending on eager or lazy open behavior or POSIX unlink semantics. Signed-off-by: YangYanbin <warlock.yyb@alibaba-inc.com>

Signed-off-by: YangYanbin <warlock.yyb@alibaba-inc.com>

Document the remaining repeated footer reads across projected DataSplit streams until paimon-rust exposes a reusable scan-scoped read context. Signed-off-by: YangYanbin <warlock.yyb@alibaba-inc.com>

Project the parent read schema by needed_columns and preserve the requested column order for both direct-file and data-split readers. Cover reordered subset projections for Parquet, Vortex, and MOR data splits. Signed-off-by: YangYanbin <warlock.yyb@alibaba-inc.com>

yanbinyang · 2026-08-12T07:57:01Z

Updated in 31c3bf1. read_schema is now treated as the parent schema and projected by needed_columns in the requested order. Added coverage for reordered subset projections across Parquet/Vortex direct-file and data-split paths, including both get_chunk and take.

sre-ci-robot added the do-not-merge/work-in-progress label Aug 6, 2026

sre-ci-robot requested review from shaoting-huang and tedxu August 6, 2026 04:16

sre-ci-robot added the size/XXL label Aug 6, 2026

yanbinyang mentioned this pull request Aug 6, 2026

[Feature]: Add Apache Paimon table format support #593

Open

3 tasks

yanbinyang marked this pull request as ready for review August 6, 2026 05:04

sre-ci-robot removed the do-not-merge/work-in-progress label Aug 6, 2026

yanbinyang mentioned this pull request Aug 6, 2026

feat: support Apache Paimon external collections milvus-io/milvus#51897

Draft

jiaqizho reviewed Aug 7, 2026

View reviewed changes

yhmo reviewed Aug 7, 2026

View reviewed changes

yanbinyang changed the title ~~feat: support Apache Paimon data-split reads~~ feat: support Apache Paimon non-direct-file reads Aug 7, 2026

yanbinyang added 4 commits August 7, 2026 15:48

fix: validate Paimon data-split descriptors

e8758f0

Reject persisted data-split descriptors with a zero row count because planning never emits them and they otherwise bypass stream count validation. Signed-off-by: YangYanbin <warlock.yyb@alibaba-inc.com>

yanbinyang force-pushed the yanbin-dev-3.0-support-paimon-data-split branch 2 times, most recently from 123a30c to e8758f0 Compare August 7, 2026 16:32

jiaqizho reviewed Aug 10, 2026

View reviewed changes

liliu-z reviewed Aug 10, 2026

View reviewed changes

fix: validate loon numeric arguments

e06f111

Parse demo-table row and dimension values without throwing on malformed input, and return the existing command error path instead. Signed-off-by: YangYanbin <warlock.yyb@alibaba-inc.com>

yanbinyang added 9 commits August 11, 2026 10:29

refactor: reorder Paimon reader declarations

f287288

Signed-off-by: YangYanbin <warlock.yyb@alibaba-inc.com>

refactor: share Paimon data-split readers directly

400e055

Signed-off-by: YangYanbin <warlock.yyb@alibaba-inc.com>

fix: preserve Paimon data-split output schemas

6343a57

Rebind projected batches to the requested Arrow schema when column names and types match, and reject incompatible schemas before they reach take or range consumers. Signed-off-by: YangYanbin <warlock.yyb@alibaba-inc.com>

fix: reuse Paimon streams for sequential chunk reads

81e4d6f

Keep a reader-local forward cursor for get_chunk, and reopen it only for backward reads or failures. Batch and range operations continue to use independent streams. Signed-off-by: YangYanbin <warlock.yyb@alibaba-inc.com>

refactor: pass Paimon projections as owned strings

19d30d4

Signed-off-by: YangYanbin <warlock.yyb@alibaba-inc.com>

docs: note Paimon projected stream metadata reuse

41f4193

Document the remaining repeated footer reads across projected DataSplit streams until paimon-rust exposes a reusable scan-scoped read context. Signed-off-by: YangYanbin <warlock.yyb@alibaba-inc.com>

		@@ -34,7 +37,38 @@
		namespace milvus_storage::paimon {
		namespace {

Uh oh!

Conversation

yanbinyang commented Aug 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What changed

Routing

Test plan

Uh oh!

sre-ci-robot commented Aug 6, 2026

Uh oh!

codecov Bot commented Aug 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jiaqizho commented Aug 7, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

yanbinyang commented Aug 7, 2026

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

yanbinyang Aug 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

yanbinyang commented Aug 6, 2026 •

edited

Loading

codecov Bot commented Aug 6, 2026 •

edited

Loading

yanbinyang Aug 11, 2026 •

edited

Loading

yanbinyang Aug 11, 2026 •

edited

Loading

yanbinyang Aug 11, 2026 •

edited

Loading