amber-lang 2.9.0__tar.gz

amber_lang-2.9.0/CHANGELOG.md

@@ -0,0 +1,53 @@
+# Changelog
+
+## 2.9.0
+
+- **`pip install amber-lang`**: it's on PyPI now, still `import amber`. Wheels for Linux (x86_64, aarch64) and
+  Apple silicon on Python 3.10 to 3.14, and the sdist carries the engine's source, so it builds anywhere with a C
+  compiler. Releases go out from a tag, through trusted publishing.
+- **Amber 2.9's temporals**: dates, months, timestamps, timespans, times, seconds and minutes come back as
+  `datetime64`/`timedelta64` with the epoch moved right (Amber counts from 2000, NumPy from 1970) and nulls as `NaT`.
+  Timespans stay zero-copy. `datetime`, `date`, `time`, `timedelta`, NumPy and pandas datetimes go in as the real
+  thing, aware ones as UTC. They used to go in as plain nanosecond numbers.
+- **Arrow both ways, properly**: `from_arrow` takes sliced and chunked tables, readers, dictionaries, decimals, every
+  int width, bools with nulls and every temporal type. Things it can't hold (nested, binary, a uint64 past the int64
+  range) say so, they used to read garbage. `to_arrow` gives temporal columns their Arrow types and bytes as `int8`.
+- **pandas**: nullable `Int64`/`Float64`/`boolean`, categoricals and tz-aware columns import, duplicate column names
+  are an error, and `to_pandas(categorical=True)` gives symbols as a `Categorical` straight from the engine's ids.
+- **Sturdier values**: `release()`, `equals()`, pickling, `symbol_codes()`. A released value raises instead of
+  crashing, views can't be written through, and every NumPy view keeps its value alive.
+- **Conversions in**: big-endian, strided, `float16` and unsigned buffers, NumPy scalars, and ints past int64 are
+  an `OverflowError`, they used to wrap.
+- `am.q` takes a query over several lines, `am.load` loads `.q` scripts and databases as `\l` does, and
+  `Table` got `head`, `tail`, `dtypes`, `to_numpy` and `from_polars`.
+- 255 tests, up from 30: Hypothesis round trips, leak and thread checks, and the CLI.
+
+## 1.9.6 — first release
+
+The Amber engine in your Python process, with zero-copy NumPy, pandas and Arrow interop.
+
+Tracks Amber engine release 1.9.6, which added the `libamber.so` dynamic C API
+this package is built on. Everything here lives outside the engine repository and
+reaches it only through that documented seam.
+
+- **Zero-copy out.** An Amber vector is exposed through PEP 3118, so
+  `numpy.asarray(v)`, `memoryview(v)` and `pyarrow.py_buffer(v)` all alias the
+  engine's own column payload and all keep it alive for as long as they exist.
+  One implementation, no NumPy or Arrow header anywhere in the build.
+- **Dynamic dispatch.** `am.gentq(10_000_000)` resolves whatever `gentq` is in
+  the engine's namespace at the moment you touch it, dotted names included.
+  Nothing in the package enumerates Amber's vocabulary, so it cannot fall behind.
+- **Bare qSQL.** `am.q("select vwap:wavg[sz;px] by sym from trades")` — the
+  rewrite the REPL applies to interactive lines, applied where the REPL cannot
+  reach.
+- **Self-contained wheels.** `libamber.so` and the eight `.k` standard-library
+  modules ship inside the wheel, so `pip install amber` gives a working engine
+  rather than a working binding that cannot find one.
+- **Keyed results** (`select … by sym`) arrive as a table with the grouping
+  columns recorded, and `to_pandas()` puts them on the index.
+- 30 tests, including pointer-equality proofs that the NumPy array and the Arrow
+  buffer are the engine's own memory.
+
+### Licence
+
+GNU AGPLv3, matching the engine.