Testing - history

How the test suite and its rules came about. What the suite is and how to run it is in Testing. Newest first; undated entries last.

2026-10-02 - the suite from 35 to 14.5 minutes

The whole suite - some 730 tests, a broker (or several) started and stopped for each - took about 35 minutes on the machine of doc/performance.md. Since 2026-10-02 it takes 14.5 (732 tests, mongoc, 868 s). A test of ~2.4 s spent its time like this:

	before	after
stopping the broker	~1.2 s	~0.15 s with mongoc (its own close), ~0.01 s with corDB
waiting for its port at start-up	up to 0.2 s	up to 0.02 s
dropping the database	~0.3 s (mongosh)	~0.008 s (corMongoDrop)

• The broker took a second to exit. On SIGTERM it joins its periodic-work thread, which slept for a second at a time; the stop now wakes it (corNgsild #56). And the helper polled for the exit with a pgrep (~30 ms) every 0.1 s - it now watches the pid it started, 10 ms at a time, without forking (#208).
• The port was polled every 0.2 s - 20 ms now (corTest#10).
• Every drop started mongosh, a JavaScript runtime, ~0.3 s - and the suite drops 1 466 times. corTools' corMongoDrop is one libmongoc connection and a command or two: ~8 ms (20 -> 14.5 minutes). mongosh stays the fallback where corMongoDrop was not built (no libmongoc on that machine).

The functests over cor:// found four server bugs

Running the suite with COR_TRANSPORT=cor is how four cor:// server bugs were found (corRest#22).

Why functional tests never read the log

A dds.so with its traces compiled out (a perf build) timed out every DDS test while the broker was working fine: the tests waited on trace lines that build never wrote. Since then a test waits on, and asserts on, what the API says.

The first real ROS 2 publisher

The first run of bridge_dds_ros2_roundtrip, against a publisher that is not ours, showed that the Enabler hands over an envelope rather than the sample.

Coverage figures in one place

The coverage figures used to be repeated in testing.md as well as coverage.md, and the two copies drifted across two different measurement regimes — testing.md went on quoting a denominator that covered coraine/src alone for a fortnight after the measurement widened to include the three libraries, so the repository stated two different coverage figures depending on which page you opened. The figures now live in coverage.md only.