[CICP-5405] Fix ZkBucketDataAccessor data loss when compressed size is an exact multiple of bucket size

[thestreak101]

Problem

ZkBucketDataAccessor computed the last bucket's length with a modulo:

compressedRecord.length % _bucketSize

When the compressed payload's size is an exact multiple of the bucket size, that expression is 0, so the final bucket is written empty.

On read, the reconstructed array keeps a zeroed tail, GZipCompressionUtil.uncompress fails, and the accessor throws HelixException("Failed to decompress path").

The write has already destroyed the bytes by then, so this is not recoverable by the reader — the record is permanently unreadable. This affects any bucketized write (WAGED baseline / best-possible assignments) that happens to compress to a multiple of the bucket size.

Fix

1. Write path — derive the last bucket from the remaining bytes instead of a modulo:

Arrays.copyOfRange(compressedRecord, ptr, compressedRecord.length)

2. Read path (a second, latent bug found while fixing the first) — numBuckets was computed from the reader's _bucketSize, while the subsequent arraycopy used the writer's bucketSize read from the metadata node. Any change to the configured bucket size between write and read produced a truncated or over-read payload. Both now use the metadata bucketSize, which is the value the data was actually written with.

3. Unified the copy loop to Math.min(bucketSize, dataSize - copyPtr) so the final short bucket needs no special case.

Tests

Added to TestZkBucketDataAccessor:

Test	Covers
testExactMultipleOfBucketSizeRoundTrip	the reported failure
testBucketBoundarySizes	sizes around N*bucketSize ± 1
testReadWithDifferentBucketSizeThanWrite	bug (2) above
testSingleBucketRoundTrip	small-payload path

A payload's compressed size is not predictable from its input size, so the helpers search for an input whose compressed size lands exactly on a bucket boundary rather than assuming one.

Verification

• mvn -pl helix-core -am test-compile passes.
• The new tests require a live ZK ensemble and were not executed locally — relying on CI.

Risk

Low. The change is confined to bucket-boundary arithmetic; the non-multiple case produces byte-identical output to before.

🤖 Drafted with GitHub Copilot CLI. Filed from an audit of CICP Helix bug tickets — please review carefully before merging.