Repository navigation
NO_TOKEN creates zombie entries in _reader map before validity check #84
Description
Activity
- added a commit that references this issue
on Mar 13, 2026 Resolved by #108. All three removal paths (
removeGenericReader,remove_registration,remove_reader_helper) now scan every bucket, including NO_TOKEN, and erase buckets that become empty.On
mainthe failure is sharper than described here. A reader removed while disconnected stays in the NO_TOKEN bucket,reconnect()moves it into a live bucket, and the removed callback keeps receiving data. Verified: 1 entry reached the removed callback onmain, and 0 on #108, #109 and #111.The NO_TOKEN bucket itself is intentional: it is the pending queue that
reconnect()drains. Readers registered while disconnected still start after reconnect on every tree.I've suggested adding
Fixes #84to #108, which targetsmain.
Description
In
RedisAdapter.cpp:274-280,add_reader_helperinserts into_reader[token]before checking iftoken == NO_TOKEN:While the reconnect logic (line 426) does handle
NO_TOKENentries, this ordering is fragile and creates map entries that may never be cleaned up if reconnect doesn't fire.Severity
High
Suggested Fix
Either check
token == NO_TOKENbefore accessing the map, or add a comment documenting this as intentional behavior for the reconnect path.