v2.38.0

v2.38.0 - 2026-09-08

🚀 Features

• feat: implement Dragonfly's CL.THROTTLE rate-limiting command, ported from Dragonfly's own code so
  the quirks match — millisecond-ceiling time fields, remaining truncating towards zero, and an
  over-sized request reporting retry_after = -1 without writing the key (#536)
• feat(dragonfly): match its errors, validation and limits — backticked unknown-command errors,
  unsigned numkeys/COUNT, SCAN COUNT 0 falling back to the default batch size, SMOVE checking
  its destination type up front, no LCS at all, a relative expiry past 2**28-1 seconds clamped
  while an absolute one is rejected, hash-field TTLs capped at 2**26 seconds, and strings capped at
  256MB rather than Redis' 512MB (#555, #556)
• feat(dragonfly): match the replies it sends — doubles from INCRBYFLOAT/HINCRBYFLOAT, empty arrays
  from a timed-out blocking pop and from an XREAD/XREADGROUP that matched nothing, ZPOPMIN/
  ZPOPMAX always paired under RESP3, flat WITHSCORES from ZUNION/ZINTER, a null
  first-entry/last-entry for an empty stream, -1 as XINFO GROUPS' unknown-lag sentinel,
  XPENDING looking up the key before the group, and one namespace for shard and plain pub/sub
  channels (#557, #558, #559)
• feat(dragonfly): match its command behaviour — SORT without -> hash-field patterns and without
  Redis' weight tie-break, COPY without a DB option, GEORADIUS/GEOSEARCH in geohash order
  unless sorted, a watched key dirtied by any write, Lua 5.4 scripting (non-integral numbers as
  doubles, Redis 6-style error wrapping, admin commands refused inside scripts), and its own JSON
  quirks (legacy paths wrapped like JSONPath, no filter expressions) (#560, #561, #562, #563)

🐛 Bug Fixes

• fix: redis.call in a Lua script now returns RESP2 shapes whatever protocol the client negotiated,
  matching real Redis, where a script must opt into RESP3. redis.setresp(2)/redis.setresp(3) are
  now supported and reset to RESP2 for each run. Dragonfly has no redis.setresp, so fakeredis does
  not expose one when emulating it, and a script returning {double=...} there answers with an empty
  array the way the real server does (#543)
• fix: XCLAIM/XAUTOCLAIM now move an entry's pending count to the claiming consumer instead of
  leaving it on the previous owner, so a later XACK no longer drives that count negative (#548)
• fix: XACK no longer raises a raw Python KeyError when a pending entry has no live consumer, and
  XGROUP DELCONSUMER now removes the entries that consumer owned, returning how many it dropped
  (#549)
• fix: XCLAIM now honours RETRYCOUNT, and XCLAIM/XAUTOCLAIM honour JUSTID, when updating an
  entry's delivery counter. RETRYCOUNT takes precedence over JUSTID, and a negative RETRYCOUNT
  means "not given" (#544)
• fix: cancelling a blocking async command (BRPOP, BLPOP, XREAD BLOCK 0, ...) no longer leaves
  the connection unusable — it went back to the pool with its socket still paused, so every later
  command on it hung (#471)
• fix: XAUTOCLAIM now returns the cursor its next scan should start from, and 0-0 once the pending
  list is exhausted (#547)
• fix: SMOVE no longer raises WRONGTYPE for a wrongly typed destination when the source key is
  missing, BITCOUNT reports the syntax error before decoding its range arguments on a pre-7.0
  server, and GEORADIUS returns an unsorted reply in geohash order rather than insertion order
  (#553)
• fix: reap TcpFakeServer handler threads when a client disconnects (#541)

🧰 Maintenance

• docs: docs/dragonfly-support.md now writes up every Dragonfly divergence fakeredis emulates, and
  the handful it does not (#564)
• test(hypothesis): the stateful machines now run against Dragonfly in CI, with the non-deterministic
  corners kept out (skiplist-encoded sorted sets, hash-order aggregation, SAVE, short-circuiting set
  intersections) (#564)

New Contributors

• @ProchazkaDavid made their first contribution in #547
• @arththebird made their first contribution in #541

Full Changelog: v2.37.1...v2.38.0