[xe-nvdk]

Summary

Brings the Python SDK (arc-tsdb-client) up to date with the Arc server after a long gap, fixes several confirmed bugs, and adds the first real HTTP-layer test coverage. Ships as v0.2.0. Sync and async paths are both covered throughout.

New capabilities

• query_msgpack() — experimental columnar MessagePack query response with native timestamp decoding; returns the same QueryResponse shape as query() (plus per-column Arrow types). Raises ArcNotSupportedError on 501 (server built without the duckdb_arrow tag).
• Bulk import (client.import_) — CSV, Parquet, line-protocol file, and TLE. Accepts paths, file-like objects, or bytes.
• TLE streaming write (write.write_tle) + tle_stats.
• Databases management (client.databases) — list / get / list_measurements / create / delete.
• Full enterprise surface (sync + async) — rbac, governance, audit, queries (query management), cluster, compaction, tiering, backup, mqtt.
• zstd write compression — now the default (gzip/none still supported) via a Compression enum. No Content-Encoding header (server auto-detects from magic bytes).
• Automatic retries — exponential backoff + jitter for connection errors, 429 (honoring Retry-After), and 503.
• New exceptions: ArcNotSupportedError (501), ArcLicenseError (403 for a missing enterprise-license feature).

Fixes

• Polars datetime→columnar used a fragile private attribute that raised at runtime; now casts via the public polars API and normalizes time_unit.
• Pandas timestamp scaling honored only nanoseconds; now respects the column resolution — including tz-naive datetime64 columns whose dtype has no .unit (read via numpy.datetime_data). These previously wrote timestamps ~1000× too small (landing in ~1970). Caught by adversarial review after an initial incomplete fix.
• Buffered writer now runs a background flush timer (thread for sync, task for async) so idle buffers still flush on flush_interval.
• show_tables validates the database identifier before SQL interpolation.
• User-Agent derived from the package version (single source of truth); removed the dead dataframe package stub.

Testing

• First real HTTP round-trip coverage via pytest-httpx: header assembly, error mapping (incl. 501/license), on-the-wire gzip/zstd magic bytes, retry/backoff, msgpack columnar decode, imports, databases, and enterprise wrappers — sync and async.
• Opt-in live integration tests under tests/integration (gated on ARC_TEST_URL).
• 167 unit tests passing; ruff + ruff-format + mypy clean.

Review

Four independent adversarial reviewers ran over the diff (async/sync twin parity, wire-protocol fidelity vs the Go server, concurrency/retry correctness, algorithmic edge-cases). Three came back clean; the algorithmic pass surfaced the tz-naive timestamp corruption bug above, which is now fixed and covered by a 4 resolutions × {tz-naive, tz-aware} regression matrix.

🤖 Generated with Claude Code

[jallegri]

Parity and merge analysis: https://github.com/Basekick-Labs/ops/issues/40.

Source inventory finds 120 SDK operations here versus 182 explicit server routes on pinned main; 56 remaining public/operational routes need scope decisions after excluding internal protocols and equivalent aliases. Method presence is not a contract pass. PR #3 is stacked on this branch; local composition with #5 conflicts in uv.lock, while #6 composes cleanly at source level.

Fresh exact-module fixtures preserve merged PR #1's sparse-column padding in sync/async across all heads, and fail on the pre-fix control. Add that oracle to the permanent regression suite. Binary query row-cap handling and nanosecond precision need explicit acceptance coverage; no full SDK/server runtime pass is claimed.