d-led / gen_server_virtual_time / 3227e9d1f9cc21c6c59fa027f6b29b12926351e8

88%

Build Type

push

github

Committed by d-led

Commit Message

fix: close virtual-clock races, stabilise tests

The clock acknowledged every message its actors handled, so an unrelated
message (for example the ack for start_sending) could satisfy the wait for a
delivered event. The clock then advanced past the event's timestamp, and the
remaining ticks were scheduled beyond the target, silently truncating a
simulation's message counts. Deliveries to VirtualTimeGenServer actors now
carry a token and are acknowledged by that token alone.

Along the way, from the new property suites and muex:

* VirtualClock.scheduled_count/1 counted distinct due times rather than
  events, so two timers at the same instant reported one.
* Events due at the same instant fired in reverse scheduling order; real
  timers fire in creation order.
* VirtualClock.cancel_timer/2 returned :ok where TimeBackend (and
  Process.cancel_timer/1) promise the remaining time, or false.
* ActorSimulation.run/2 read actor statistics in a single pass, so a
  downstream actor could be sampled before an upstream one had forwarded to
  it. Statistics are now read until they settle.

Also: the ack watchdog is tracked in the clock state, so a stale timeout can
no longer fire against a later wait; all three compiler warnings are gone
(dead clauses); the four duplicated dispatch blocks in Actor are one helper;
and advance_loop's triplicated timeout logic is extracted.

Dependencies and CI:

* Updated all dependencies. ex_doc 0.40 requires Elixir 1.15, so the declared
  floor moves up from 1.14.
* Replaced muzak and exavier with muex: both are broken on current Elixir
  (muzak crashes in its tokenizer, exavier calls a removed ExUnit function).
* Added stream_data and thirteen property tests covering time additivity,
  event ordering, cancellation, scheduling counts and simulation counts.
* Reduced the CI matrix to Elixir 1.19/1.20 with OTP 28/29, and aligned the
  other five workflows, which still pinned 1.15/1.18 and OTP 25/27.

Test reliability:

* Replaced ... (continued)

Coverage Stats

89 of 110 new or added lines in 6 files covered. (80.91%)

2 existing lines in 1 file now uncovered.

2117 of 2445 relevant lines covered (86.58%)

5087.78 hits per line