7.1.0

Overview

7.1.0 adds native decimal and double-array column APIs and a non-blocking
completion poll, and fixes several QWP/WebSocket correctness and performance
issues. It also changes how UUID bytes and Arrow FixedSizeBinary columns are
interpreted: read "Breaking changes" before upgrading if you send or read
UUIDs as raw bytes or ingest Arrow, Polars, or NumPy data.

The C and C++ clients are distributed as source. No pre-built native binaries
are attached; use the tag's source archive.

Compatibility

• QWP over WebSocket requires QuestDB 10.0 or newer. HTTP Basic auth and
  single-endpoint reconnect work with Open Source and Enterprise; bearer-token
  and OIDC auth, multi-host failover, and durable acknowledgement require
  QuestDB Enterprise.
• Rust MSRV is 1.91.0, relaxed from 1.91.1 (#211).
• The published questdb-rs crate supports Arrow >=58, <60 and Polars
  >=0.52, <0.55. The native (FFI) library pins Arrow =59.0.0 (#195).

See doc/COMPATIBILITY.md
for the full matrix.

Breaking changes

• UUID byte order (#186). Raw UUID bytes are now in standard RFC-4122
  order, matching Python's uuid.bytes and Rust's Uuid::as_bytes(). This
  affects column_uuid() with a byte array, NumPy S16 (qwp_numpy_s16),
  bind_uuid, Arrow UUID columns, and every reader getter. Code that swapped
  bytes to match the previous order must stop doing so. The lo/hi integer
  UUID API is unchanged.
• Arrow FixedSizeBinary columns ingest as BINARY unless tagged (#186).
  A UUID column needs the arrow.uuid or questdb.column_type=uuid field tag
  (16 bytes), or ArrowColumnOverride::Uuid at flush time; LONG256 needs
  questdb.column_type=long256 (32 bytes) or ArrowColumnOverride::Long256.
  A tag whose width doesn't match the column is an error.
• Invalid Arrow column metadata is an error rather than being ignored
  (#186).
• Arrow field metadata is capped at 64 MiB per batch through the C, C++,
  and Python APIs (#195). Larger batches are rejected and nothing is sent.
  Polars Enum columns with millions of categories should be cast to
  pl.Categorical. Rust flush_polars_dataframe is not affected.
• A failed columnar batch can publish a prefix (#219). A batch of up to
  eight rows that contains an oversized row, other than the first, now sends
  the rows before it instead of being rejected whole. A store-and-forward
  sender commits those rows and reports the error as in doubt; a direct sender
  closes its connection, discarding its other uncommitted chunks, so the
  following commit fails with FailoverRetry. Check Error::in_doubt() before
  retrying the whole input.

New

• Native Chunk column APIs for Decimal64, Decimal128, Decimal256, and
  fixed-shape DOUBLE arrays, in Rust, C, and C++ (#206). Buffers are borrowed
  through flush, with validation of scales, shapes, lengths, and validity
  bitmaps.
• Sender::completed_fsn(AckLevel): a non-blocking poll of the completion
  watermark, the polling counterpart to Sender::wait (#190).
• QWP row buffers accept symbols after other columns (#203), over both UDP and
  WebSocket. ILP still requires symbols before fields.
• Polars Object columns fail with a clear error instead of ingesting
  meaningless bytes (#186).

Fixes

• Columnar batches whose rows each fit max_buf_size are no longer rejected;
  splitting continues down to single rows (#219).
• Error::in_doubt() on a failed QuestDb::flush_polars_dataframe() covers
  the whole DataFrame (#195): if any earlier batch was published, replaying the
  DataFrame could duplicate that prefix.
• Byte-aligned GEOHASH maximum values are no longer encoded as NULL (#195).
• Arrow C Data Interface inputs are validated before import: schema/array
  agreement, child counts, buffer layouts, offsets, and slice lengths (#195).
  On failure, inspect array->release to determine ownership; the error code
  alone does not.
• Closing a connection pool (questdb_db_close) could block forever on a
  lost wakeup in the rejection-event dispatcher (#189).
• In store-and-forward memory mode, a split chunk commits once at its last
  frame instead of once per frame, with a bounded deferred window (#178). The
  memory-mode default store-and-forward budget is at least five segments.
• Buffer::bookmark() on a QWP/WebSocket buffer records lengths instead of
  cloning the buffered data, so bookmarking per row is no longer quadratic
  (#191).
• QwpWsSizeHint::len() reuses its scratch allocation (#190).

Other changes

• The InvalidApiCall message returned by wait for AckLevel::Durable
  without request_durable_ack=on has new text (#190). The error code is
  unchanged.

Full changelog: 7.0.0...7.1.0