Version 2.4.0

What's Changed

New Features

• Cluster --monitor-input support: Route-then-stage replay across shards with end-to-end backpressure (#395, #399).
• Cluster --multi-key-get: Slot-aware MGET routing with a shared cache rebuilt on every topology refresh (#397).
• --transaction flag: Pins a --command rotation to a single shard in --cluster-mode; works with --pipeline > 1 (#390, #400).
• --retry-on-error pipeline: Adds --max-retries, --retry-backoff-ms, --retry-backoff-factor, --retry-on=LIST, --max-retry-queue, --failed-keys-file (#381).
• --realtime-latencies: Streaming HDR-histogram block emits per-second cur (avg: …) for ops/sec, bytes/sec, miss%, and configured percentiles (#379).
• Per-key miss tracking for --command: --command-miss-tracking={auto,off} plus --miss-rate-threshold=PCT; emits a Per-Key Misses JSON section (#374, #398).
• --connection-stage-timeout=SECS: Bounds the AUTH/HELLO/SELECT/CLUSTER SLOTS / first-steady-state phase; defaults to 30 s, 0 disables (#431).
• TLS handshake logging: New TLS connection established: protocol X, cipher Y stderr line and a top-level TLS object in mb.json (#403).

Bug Fixes

• Stack overflow in --monitor-input replay: VLA-allocated split-token buffer overflowed for large lines; now heap-allocated (#405).
• Stack smash in memcached text parser: Unbounded sscanf("%s %s …") allowed a hostile server to overflow a 256-byte stack buffer; widths now bounded (#415).
• OOB read on malformed CLUSTER SLOTS: handle_cluster_slots now validates reply shape and clamps slot bounds; malformed shards are logged and skipped (#425).
• SIGPIPE on TLS RST: signal(SIGPIPE, SIG_IGN) installed process-wide so a peer RST mid-write no longer exits 141 (#383).
• Crash handler hardening: sigaltstack on main and each worker, pre-warmed backtrace symbols, backtrace_symbols_fd, and a 5 s watchdog (#412, #416).
• Cluster reconnect SIGABRT: cluster_client::connect() is now idempotent under --cluster-mode --reconnect-on-error (#378).
• Reconnect duplicate-error thread kill: Duplicate error callbacks while a reconnect is pending no longer break the event loop (#392).
• Unbounded reconnect/retry backoff: Exponential backoff is now capped at 60 s; previously factor=2.0 + unlimited attempts could schedule reconnects ~34 years out (#433).
• --monitor-input parsing: Preserves embedded NUL bytes in the loader; handles CR-only line endings (#434).
• RESP3 push and nil: Accepts > push frames (drained out-of-band) and counts _ (nil) as a miss in arbitrary-command tracking (#435).
• CLI input validation: Rejects at parse time inputs that 2.3 silently accepted and then crashed/hung on — -c/-t/-n/--test-time negatives (#436), --run-count overflow (#429), --pipeline / --data-size / --data-size-list (#428), --command / --command-ratio (#427), --key-stddev / --key-median / G:G degenerate ranges (#430). All return a clear error and exit 2.
• Bound recursive RESP teardown: Iterative ~mbulk_size_el so deeply-nested replies can't stack-overflow on cleanup; -Werror=vla build guard (#415).
• TLS path leakage in JSON: --cert/--key/--cacert are now basename-only in mb.json (#440).
• m_reqs_generated defensive clamp: Underflow guard at three cluster_client subtraction sites (#437).
• Hits/sec /0 guard: summarize() rate computations no longer emit +Inf/nan on zero-duration runs (#436).

Build & Configuration Improvements

• OpenSSL 4.0 build: Replaces removed SSL_OP_NO_TLSv* flags with SSL_CTX_set_min/max_proto_version; rejects non-contiguous --tls-protocols selection at parse time on 4.0+ (#432, #444).

Developer Tooling & CI Improvements

• libFuzzer harnesses: Per-PR 60 s smoke fuzz across split_command_to_args, monitor_input, and redis_protocol (#421).
• Adversarial RESP mock-server fixtures: 12 deterministic payloads exercising the parser surface (#419).
• Monitor-input black-box fuzzer: Corpus + synthetic seeds, nightly + run-fuzz label trigger (#422).
• Hypothesis CLI fuzzer: Re-broadened strategy + new cli-fuzz-nightly.yml (#424, #445).
• Differential vs redis-benchmark / redis-cli: run-differential PR label, label/cron-triggered workflow (#420).
• Nightly soak suite: 7 scenarios (long-run memory, large payloads, high concurrency, connection churn, cluster turbulence, retry storm, slow network) with run-soak label (#423).
• Sanitizer stress matrix: STRESS=1 cells across ASAN/UBSan/TSAN (#418).
• Workflow permissions hardening: permissions: contents: read declared on every workflow (#380, #393).
• PARALLELISM=2 on cluster cells: Resolves Test cell 15-min timeout on cluster TLS (#402).

New Contributors

• @arpitjain099 made their first contribution (#380, #393).

Full Changelog: 2.3.1...2.4.0