GeezerVison

Resolves #70.

What

Adds Server.readlines_to_eof_timed(timeout) — a bulk-drain counterpart to readline_timed that uses the isc_info_svc_to_eof request mode (the same one Firebird's fbtracemgr uses internally). Returns the multi-line chunk delivered in one services-API round-trip, distinguishes end-of-service (isc_info_end) from transient idleness (isc_info_svc_timeout / isc_info_data_not_ready / isc_info_truncated), and lets callers drain a busy trace stream at fbtracemgr-equivalent rates.

Why

readline_timed sends bytes([SrvInfoCode.LINE]) and gets one line per call. Under sustained trace-stream load a Python consumer is throttled to roughly 480 events/sec (10k lines/sec × ~21 lines per event of typical OLTP trace output). fbtracemgr (which sends isc_info_svc_to_eof) reaches roughly 3000 events/sec on the same workload; the gap is purely the request-mode round-trip overhead. When the Python reader falls behind, events accumulate in Firebird's server-side trace buffer until the engine discards them — and the consumer never knows.

Concrete numbers from a matched-workload comparison (Firebird 5.0.3 on Windows; HammerDB TPC-C 200-call transaction mix against a pre-loaded 1-warehouse schema; identical trace.conf for both readers):

Same Python process, same python3-driver connection — the only delta between rows 2 and 3 is the request item. Full investigation write-up in #70.

Design choices

1. Additive, not destructive. readline_timed and its one-line semantics stay exactly as-is. Server.readline, stdin handling, and any third-party callers depending on per-line behavior keep their current behavior. The new method is opt-in for consumers (like trace-stream readers) that need bulk throughput.

2. Trailing-tag-aware EOF. The Firebird engine (src/jrd/svc.cpp) emits one of four trailing indicators after a to_eof response:

   • isc_info_svc_timeout (64) — server timed out, session live.
   • isc_info_data_not_ready (4) — no payload available, session live.
   • isc_info_truncated (2) — buffer truncated, more data on next call.
   • isc_info_end (1, implicit) — service has finished.

   readline_timed only checks for isc_info_svc_timeout and conflates the other three into None/EOF. The new method explicitly checks for isc_info_end as the EOF signal and treats everything else as "session live, may have more data". This was the second half of the issue — a transient idle round was prematurely terminating long-lived consumers.

3. Chunk delivery, not line splitting. readlines_to_eof_timed returns the server's multi-line buffer verbatim. Per-line consumers split with splitlines(keepends=True). The driver doesn't dictate line semantics on a path designed for trace-stream consumers who often want the raw bytes anyway (to dump verbatim to a file, the way fbtracemgr does).

Tests

Two new integration tests in tests/test_server.py, matching the existing service-test style:

• test_readlines_to_eof_timed_drain — drains database.get_statistics() via both readlines() and readlines_to_eof_timed in a loop; asserts equivalent normalized output. Confirms the bulk path captures the full service response and terminates correctly on isc_info_end.
• test_readlines_to_eof_timed_returns_timeout_when_idle — starts a trace session against an idle database and asserts the very first read returns TIMEOUT, not None. Pins the "transient idle round != EOF" contract that fixes the second half of readline_timed drops trace stream events under load (uses isc_info_svc_line per-call instead of isc_info_svc_to_eof bulk drain) #70.

I have not added a unit test that mocks the services connection because the existing tests are all integration-style against a live Firebird; happy to add mock-driven tests if the project prefers that direction.

Compatibility

• Public API addition only; no signature changes.
• Server.readline (which internally calls readline_timed via _read_output) keeps the same behavior. No callers were modified.
• Tested against the existing test_server.py suite shape; no other tests were modified.

Notes

Section heading in CHANGELOG.md added as [Unreleased] to be retitled when the maintainers tag the next release.

Happy to iterate on naming (readlines_to_eof_timed vs read_to_eof_timed vs something else), splitting semantics, or test coverage. I am also fine reshaping this as a replacement for readline_timed rather than an addition if the maintainers prefer that route, though it would be a breaking change for any callers relying on the one-line semantics.