https://github.com/petergoldstein/dalli - v5.0.6

Performance:

• Skip the cas-return flag on quiet meta_set requests (#1131)

  • In quiet mode memcached suppresses the ms response entirely, so the CAS requested by the c flag can never be read; sending it only added two bytes to every request
  • Applies to the bulk-write paths, where quiet sets are emitted: Dalli::Client#multi blocks and the pipelined setter
  • Extracted from #1130; thanks to Jianbin Chen for this contribution

• Reduce allocations in KeyRegularizer and multi-key request paths (#1120)

  • Decomposed KeyRegularizer#encode into separate needs_encoding? and encode calls so the common happy path avoids allocating an intermediate array for the two-element return value
  • Refactored multi_get/multi_set/multi_delete command generation into RequestFormatter to share its key-encoding helpers
  • Thanks to Jean Boussier for this contribution

• Reduce allocations in ResponseBuffer pipelined getk parsing (#1117)

  • process_single_getk_response was building a fresh array to return results alongside the updated offset; refactored to store the offset as the last element of the existing tokens array and pop it, saving one allocation per response
  • Also skips trailing nils in the token array
  • Thanks to Jean Boussier for this contribution

• Enable frozen string literals in RequestFormatter (#1118)

  • Frozen string literals had been inadvertently disabled; re-enabling reduces allocations by ~300,000 objects in a 10,000-iteration get_multi_cas benchmark (562 MB → 550 MB total allocated)
  • Thanks to Jean Boussier for this contribution

• Reduce allocations in ResponseProcessor#value_from_tokens (#1113)

  • token[1..].to_i was allocating a new string for every token parsed; replaced with in-place slice! followed by a token reset to avoid poisoning subsequent token comparisons
  • Saves 4 allocations per entry in get_multi_cas workloads (a hotspot for IdentityCache)
  • Thanks to Jean Boussier for this contribution

• Reduce allocations in common operation paths (#1111)

  • Use Symbol#name over Symbol#to_s to return a frozen string without allocation
  • Skip trace attribute hash construction when OpenTelemetry instrumentation is disabled
  • Use argument forwarding (...) in Client#perform and Threadsafe#request to avoid splat array allocation
  • Use match? in KeyRegularizer#encode to avoid MatchData object allocation
  • Reduces objects allocated by ~26% and memory by ~6% for a simple get workload
  • Thanks to Jean Boussier for this contribution

• Fix pathological memory behavior in ResponseBuffer (#1114)

  • compact_if_needed was intended to reclaim memory by slicing off consumed bytes, but buffer.byteslice(@offset..) on an unfrozen string causes Ruby to allocate a hidden third string as the copy-on-write owner rather than freeing the original
  • Redesigns buffer management to pass reusable buffer objects directly to read/read_nonblock, avoiding reallocation on each response read
  • Reduces allocations from ~2.38 GB to ~649 MB in a get_multi_cas benchmark over 10,000 iterations
  • Accompanied by new unit tests for ResponseBuffer (#1115)
  • Thanks to Jean Boussier for this contribution

Features:

• delete_multi now returns the number of keys found and deleted (#1126)
  • Previously the return value was unspecified; callers (e.g. Rails, see rails/rails#58071) had no way to tell how many keys were actually removed
  • The count is derived from the meta protocol's quiet-mode delete responses with no extra round-trips: successful deletes are suppressed while misses report NF, so any response received before the terminator is a key that was not deleted
  • The single-server fast path now shares the pipelined path's bounded retry on transient (RetryableNetworkError) network errors, so both paths behave consistently; the returned count is best-effort and may under-report if a network error triggers a retry, since keys deleted before the error are not recounted
  • Thanks to Iliana Hadzhiatanasova for this contribution

Bug Fixes:

• Raise instead of returning a truncated value when the peer closes mid-response (#1135)

  • IO#read(count) on a blocking socket accumulates across TCP chunks and hands back a shorter buffer (or nil) in only one case: the stream hit EOF. That short buffer was passed through as the response body, so a memcached restart, proxy drop, or load balancer timeout partway through a response could surface a truncated but still decodable value to the caller, indistinguishable from a real one
  • A short read is now treated as the premature EOF it is, raising and closing the dirty socket so the request is retried on a fresh connection
  • CRuby only; the JRuby path already used Socket#readfull, which enforces the same contract
  • Extracted from #1130; thanks to Ian Ker-Seymer for the original fix and Jianbin Chen for the port

• Tear down the connection when a non-StandardError aborts a request (#1136)

  • Async::Stop and Thread#kill descend from Exception rather than StandardError, so the rescue clauses in Protocol::Base#request never saw them; a scheduler cancelling a fiber parked on a response read skipped close entirely, leaving the connection marked as having a request in progress with partial response bytes still unread on the wire, and returning that half-used client to the pool under connection_pool
  • Protocol::Base#request now closes in an ensure unless the request ran to completion, and ConnectionManager#close performs its state cleanup in an ensure so a second cancellation landing inside @sock.close cannot leave the socket non-nil with the request still marked in progress
  • Dalli::DalliError and Dalli::MarshalError now close the connection at the point of failure rather than at the start of the next request; those paths already left the request in progress and ConnectionManager#confirm_ready! closed on the next call, so this changes when the close happens rather than adding one
  • Extracted from #1130; thanks to Dan Mayer for the original fix and Jianbin Chen for the port

• Fix ResponseBuffer compaction logic (#1119)

  • COMPACT_THRESHOLD was removed in #1116 as apparently unused, but the constant was referenced by the compaction guard; its absence silently disabled buffer compaction
  • Restored the constant, corrected the compaction condition, and improved the buffer-shrinking implementation to use String#bytesplice (backed by memmove) for true in-place compaction
  • Adds targeted tests covering the compaction threshold and shrink behavior
  • Thanks to Jean Boussier for this contribution

Maintenance:

• Scope StrictWarnings to Dalli's own source (#1134)

  • The test suite runs under -w and prepends a hook to Warning.singleton_class that turns warnings into failures, but that hook is global: a warning emitted while loading any third-party gem aborted the whole suite before a single test ran
  • json 2.21.2's pure-Ruby generator (used on JRuby, where the C extension is unavailable) warns method redefined; discarding old to_hash at require time, which took the jruby-10 CI job red with no change to Dalli
  • Warnings are now attributed to a source file and only raise for lib/ and test/; attribution prefers the location Ruby embeds in the message, since the stack at that point describes the require chain rather than the offending code
  • Portable attribution of Kernel#warn callers also required walking the stack rather than indexing it (Ruby 3.3/3.4 push an <internal:warning> frame that 4.0 does not), skipping RubyGems' Kernel#warn shim (active on JRuby but not CRuby), and resolving relative backtrace paths

• Make raw and namespace fast path tests actually use those options (#1129)

  • Followup to #1127: the raw and namespace variants passed those options to the helper that starts memcached, which configures the client the tests then discarded, so neither option was ever exercised
  • Passes the options to the client under test and adds assertions that fail if they are absent
  • Thanks to Iliana Hadzhiatanasova for this contribution

• Benchmark set_multi and add a delete_multi target (#1132)

  • Enables the two set_multi reports that were commented out pending the arrival of set_multi, resolving the accompanying TODO
  • Adds a delete_multi target comparing the pipelined path against N single deletes
  • Extracted from #1130; thanks to Jianbin Chen for this contribution

• Bump CI memcached to 1.6.41 and run benchmarks on pull requests (#1133)

  • The tests workflow moves from 1.6.40 to 1.6.41; the benchmarks and profile workflows had drifted back on 1.6.23
  • Extracted from #1130; thanks to Jianbin Chen for this contribution

• Disable RuboCop metrics cops (#1128)

  • Thanks to Jean Boussier for this contribution

• Remove PIDCache module (#1125)

  • Process.pid is cached natively by Ruby 3.3+ (via https://bugs.ruby-lang.org/issues/19443), making the manual cache unnecessary now that Dalli requires Ruby 3.3+
  • Thanks to Jean Boussier for this contribution

• Remove unused COMPACT_THRESHOLD constant from ResponseBuffer (#1116)

  • Followup cleanup after the buffer management redesign in #1114
  • Note: subsequently found to be in use; restored and corrected in #1119
  • Thanks to Jean Boussier for this contribution

• Use String#byteindex instead of String#index when searching for the response terminator in getk_response_from_buffer (#1112)

  • The result feeds directly into byteslice; byteindex makes the intent explicit, though both return the same value since the buffer encoding is always BINARY
  • Thanks to Jean Boussier for this contribution

• Make single-server fast path tests actually exercise the fast path (#1127)

  • The batch-operation tests built clients through a helper that registers two address aliases for the same memcached process, so every client had a 2-server ring and the tests always ran through the pipelined path instead of the single-server fast path
  • Adds a single_server_client test helper that builds a client with a single address, and uses it in the affected get_multi, set_multi, and delete_multi tests
  • Thanks to Iliana Hadzhiatanasova for this contribution

- Ruby
Published by petergoldstein 15 days ago