Repository navigation
feat(java): add chunk and edge-range primitives - #968
Conversation
|
Hi @SemyonSinchenko — this is a deliberately small, independent Java core slice (419 lines, 6 focused tests) after #960/#962. It provides only the long-safe half-open chunk/range value layer; metadata URI resolution and IO stay out of this PR. I would appreciate your review when convenient. |
|
Updated this PR into the complete ordered-adjacency core vertical (now 848 additions): validated offset chunks, metadata URI resolution, and the canonical LDBC cross-chunk fixture. It remains independent of #961 by keeping Parquet reads out of core. |
|
cc @SemyonSinchenko @SYaoJun — green CI, would appreciate a look when you have time. |
eb6f816 to
a959411
Compare
|
@SemyonSinchenko @SYaoJun could you take a look when you have a moment? This one is rebased on current Note that the I can follow up with a separate PR widening |
|
Update on the CI gap I mentioned above: the fix landed in #966 rather than as a separate PR, since @SYaoJun raised the same point there. The workflow now builds
|
Provide the storage-free arithmetic layer needed by future ordered adjacency readers without coupling it to GraphAr metadata or physical IO. Relates to apache#967. Rejected: resolving metadata layouts or reading offsets in this value-only slice.
Complete the core-only ordered-adjacency contract: validate offset chunks, select exact half-open edge chunk ranges, and delegate URI resolution to immutable GraphAr metadata. Relates to apache#967. Rejected: reading Parquet offsets in graphar-core; io-parquet remains behind the open apache#961 API change.
The workflow's path filters already list maven-projects/core/**, so a change under core triggers the job, but SDK_MODULES left the module out and the reactor never compiled or tested it. A green run therefore proved nothing about core. Add core to the module list so the same job that guards storage and io-api guards it too. The fixture-backed test skips itself when the canonical testing dataset is absent, so the job stays green without the submodule.
a959411 to
a551002
Compare
|
Rebased on I also pushed the answer to the CI question I raised earlier, rather than leaving it open: the workflow already watches Local verification against |
|
I left a few minor comments. ~LGTM overall |
An ordered-layout resolve took any OffsetChunk and indexed it with an offset computed from a different vertex chunk, silently returning a plausible but wrong edge range. The chunk now carries the vertex chunk it was read from, and resolve rejects a mismatch before using it. Chunk arithmetic in EdgeRange.edgeChunks now delegates to ChunkMath instead of repeating the division, and ChunkRange, EdgeRange and OffsetLocation gained value equality plus readable toString so tests can assert whole values rather than field by field.
Reason for this PR
Implements #967 (and relates to #947/#944): the complete dependency-light Java core contract for resolving an ordered GraphAr adjacency range.
What changes are included in this PR?
graphar-corewith long-safe chunk arithmetic and half-open edge/chunk range value types.EdgeInfoordered-by-source/destination metadata into the exact offset location, edge range, edge-count URI, and selected adjacency chunk URIs.[1008, 1061)at chunk size1024.This PR delegates URI construction to
EdgeInfo; it deliberately does not read Parquet offsets or depend on storage/Parquet. That physical reader belongs after the open #961 IO API change.Are these changes tested?
Yes.
The reactor passes 8 core tests plus the
graphar-infodependency suite.pre-commit run --files ...andgit diff --checkalso pass.Are there any user-facing changes?
Yes. Java consumers can now resolve ordered-adjacency metadata and validated offset ranges to exact GraphAr chunk URIs without introducing physical IO dependencies.
Checklist
make cpplint(if this PR contains C++ changes)pre-commit