[sentry-junior]

Fixes #3038

SessionFlusher#add_session read and wrote @pending_aggregates in two lookups without holding @mutex. Only the flush swap in #pending_envelope took the lock, so the lock didn't stop anything. If the swap landed between the two lookups, the second one returned nil and raised NoMethodError: undefined method '[]' for nil from Hub#end_session, which runs in the Rack middleware's ensure. A swap at other points dropped the session silently.

add_session now holds @mutex around both lookups, which matches how TelemetryEventBuffer locks both sides. The new spec starts the flush swap on another thread while add_session is mid-update. It fails on master with the NoMethodError and passes with this change.

Proof: reproduction in getsentry/repro#65 (sentry-ruby/3038/repro.rb). It runs a deterministic interleaving plus a stress run with no hooks: one thread calls add_session while another runs the real swap.

Before (7.2.0, Ruby 4.0.9):

== 1. deterministic: swap lands between the two lookups in add_session
BUG: add_session raised NoMethodError: undefined method '[]' for nil
     at .../sentry-ruby-7.2.0/lib/sentry/session_flusher.rb:33:in 'Sentry::SessionFlusher#add_session'

== 2. stress: real add_session thread vs. real flush-swap thread (5.0s)
sessions added without error: 3144184
add_session crashes (NoMethodError): 7
sessions missing from flushed envelopes: 12

RESULT: FAIL - bug reproduced (2/2 scenarios)

After (this branch, Ruby 4.0.9):

== 1. deterministic: swap lands between the two lookups in add_session
OK: add_session did not raise and the session was flushed

== 2. stress: real add_session thread vs. real flush-swap thread (5.0s)
sessions added without error: 2798919
add_session crashes (NoMethodError): 0
sessions missing from flushed envelopes: 0

RESULT: PASS - SessionFlusher is safe against the flush swap

With RUBY_THREAD_TIMESLICE=1 STRESS_SECONDS=15, which makes MRI switch threads far more often, 7.2.0 had 286 crashes and 225 lost sessions out of about 8.6M. This branch had 0 and 0 out of about 8.4M. Ruby 3.2.11 shows the same before/after result.

via @dingsdax.

--

View Session [Sentry]