Dgraph v25.2.0 Released

[matthewmcneely]

We're pleased to announce the release of Dgraph v25.2.0. This is a binary compatible release with v25 and v24.

Those customers using vectors for similarity search should read the two caveats below with #9562 and #9559. Information regarding the rationale for changes in these PRs can be found in their respective running commentaries. Comments in #9559 include results from comprehensive BEIR SciFact information retrieval benchmarks.

Special thanks to @joelamming for their contributions and support in improving the capabilities of our vector similarity search.

Also of note, via #9568, our Docker images no longer have Critical or High CVEs.

• Added

• GraphQL

  • feat(graphql): Add support for ef and distance_threshold in generated GraphQL queries for
    similarity search (feat(graphql): Add support for ef and distance_threshold in generated GraphQL queries for similarity search #9562).

    WARNING In feat(graphql): Add support for ef and distance_threshold in generated GraphQL queries for similarity search #9562, the computed distances (automatically returned in the vector_distance
    field) for cosine and dot product metrics are no longer normalized to [0, 1], but instead
    return the raw distance computed by the metric.

• Vector

  • feat(vector/hnsw): add per-query ef and distance_threshold to similar_to, fix early
    termination (feat(vector/hnsw): add per‑query ef and distance_threshold to similar_to, fix early termination #9514)

• Fixed

• Core

  • fix(txn): for lossy indexes, change comparison function to first check the txn cache (fix(txn): for lossy indexes, change comparison function to first check the txn cache #9567)
  • fix(cmd): store correct CA verification status (fix(cmd): store correct CA verification status #9554)
  • fix(zero): make zero shutdown cleanly (fix(zero): make zero shutdown cleanly #9525)

• Vector

  • fix(vector): Fix similarity-based HNSW search for cosine and dot product metrics (fix(vector): Fix similarity-based HNSW search for cosine and dot product metrics #9559)

    WARNING To benefit from the fix in fix(vector): Fix similarity-based HNSW search for cosine and dot product metrics #9559, you must reindex vector predicates that use
    cosine or dot product metrics.

• Chore

  • docs: fix typos in comments (docs: fix typos in comments #9569)
  • chore(ci): update go toolchain version to v1.25.6 (chore(ci): update go toolchain version #9568)
  • fix(ci): update trunk go runtime to match go.mod version (fix(ci): update trunk go runtime to match go.mod version #9575)
  • chore(test): replace deprecated docker struct types in testing harness (chore(test): replace deprecated docker struct types in testing harness #9549)
  • chore(test): unskip previously skipped tests (chore(test): unskip previously skipped tests #9537)
  • chore(core): use Tick() instead of NewTicker() (chore(core): use Tick() instead of NewTicker() #9548)

New Contributors

• @mlwelles made their first contribution in fix(ci): update trunk go runtime to match go.mod version #9575
• @NAM-MAN made their first contribution in docs: fix typos in comments #9569

Full Changelog: v25.1.0...v25.2.0

---

This discussion was created from the release v25.2.0.

[rahst12]

Excited about the new version!

Can you go into a little more depth explaining this Core fix and when the scenarios occured?

• fix(txn): for lossy indexes, change comparison function to first check the txn cache (fix(txn): for lossy indexes, change comparison function to first check the txn cache #9567)

Thanks

[matthewmcneely]

Sure, there was a long-standing bug in which predicates that were mutated in a transaction would not be found in the same txn with a comparison function if the predicate was indexed with a lossy tokenizer. Lossy indexes are primary those that either reduce precision like the date(hour) index (where the bug first was spotted) or apply text processing like term or fulltext string indexes. Geo, bigfloat, vector are also lossy indexes.

Check out the issue report (#9556), @kishenpateldotwork did a great job writing up the bug. Also, this regression test that was added to validate the fix illustrates the problem -- this test fails on 25.1