sanadb 0.1.2__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -0,0 +1,6 @@
1
+ /target
2
+ /data
3
+ /python/.venv
4
+ /python/target
5
+ *.so
6
+ __pycache__/
@@ -0,0 +1,68 @@
1
+ # Changelog
2
+
3
+ All notable changes to sanadb. The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/)
4
+ and versions follow [Semantic Versioning](https://semver.org/). Before 1.0,
5
+ minor versions may change on-disk formats; when they do, this file says so and
6
+ the new version migrates old stores on open.
7
+
8
+ ## [Unreleased]
9
+
10
+ ## [0.1.2] - 2026-11-15
11
+
12
+ First version with Python wheels on PyPI: the 0.1.1 release published the
13
+ crate but its macOS Intel wheel job sat on a retired runner label and PyPI
14
+ never ran. Intel macOS wheels now build on `macos-15-intel`.
15
+
16
+ ### Changed
17
+ - Log format is now `SANADB02`: frames are padded to 4 bytes so vectors sit
18
+ at aligned offsets. The delta index reads vectors from the memory-mapped log
19
+ instead of holding copies in RAM, so bulk-loading N vectors no longer needs
20
+ N vectors of memory before the first checkpoint.
21
+
22
+ ### Added
23
+ - Forward-compatibility rules, modelled on SQLite's read/write versions. Both
24
+ file headers carry a reader level (currently 1). A build opens any file
25
+ whose level it supports and skips record kinds it does not know, so a future
26
+ version can add features without locking older readers out. A file whose
27
+ level is higher than the build supports fails with
28
+ `Error::NeedsNewerVersion` instead of being misread.
29
+
30
+ ### Upgrading from 0.1.1
31
+ - Stores written by 0.1.1 (`SANADB01`) are converted automatically the first
32
+ time they are opened. The conversion rewrites the log through a temp file
33
+ and rename; a crash during it leaves the old file intact. The index sidecar
34
+ is unchanged. There is no downgrade path: a converted store will not open
35
+ with 0.1.1.
36
+
37
+ ## [0.1.1] - 2026-11-15
38
+
39
+ First published release, on crates.io and PyPI. 0.1.0 was tagged but never
40
+ published.
41
+
42
+ ### Added
43
+ - Write-ahead log with CRC-framed records, torn-tail truncation, `sync()` and
44
+ `sync_every_write`.
45
+ - Memory-mapped Vamana (DiskANN-style) graph index built at checkpoint, with
46
+ product-quantized codes in RAM for the walk and exact re-ranking from disk.
47
+ - Parallel graph build; sharded build above `shard_size` (default 1M) over a
48
+ memory-mapped snapshot.
49
+ - Beam-width expansion with `madvise(WILLNEED)` prefetch on Unix
50
+ (`Options::io_width`, default 4).
51
+ - Exact in-RAM delta for records written since the last checkpoint, with
52
+ tombstones for deleted or overwritten graph records.
53
+ - Non-blocking checkpoint: readers and writers continue during the build.
54
+ - Metadata filters with adaptive beam widening for selective filters.
55
+ - Metrics: L2, dot product, cosine (normalises on insert and query).
56
+ - Rejection of NaN and infinite vector components.
57
+ - Thread-safe `Db` (`Send + Sync`), exclusive file lock per open store.
58
+ - Python bindings (`pip install sanadb`) with numpy input, dict or callable
59
+ filters, and the GIL released during work.
60
+ - Examples reproducing the SIFT1M and SIFT1B-prefix measurements in the README.
61
+
62
+ ### Known limits
63
+ - One process per store. No replication, encryption, typed metadata or range
64
+ filters. Formats may change before 1.0 (with migration).
65
+
66
+ [Unreleased]: https://github.com/RohithOrg/sanadb/compare/v0.1.2...HEAD
67
+ [0.1.2]: https://github.com/RohithOrg/sanadb/compare/v0.1.1...v0.1.2
68
+ [0.1.1]: https://github.com/RohithOrg/sanadb/releases/tag/v0.1.1