Dev

.gitignore

@@ -33,7 +33,16 @@ Thumbs.db
 PLAN.md
 NOTES.md
 
-scripts/
+# Local scratch scripts stay out, but the ones the README and the docs site link
+# to have to be tracked or those links 404.
+#
+# `scripts/*` rather than `scripts/`: git cannot re-include a file whose parent
+# directory is excluded, so a bare `scripts/` would make the negations below
+# silently do nothing.
+scripts/*
+!scripts/bump-version.sh
+!scripts/benchmark.sh
+!scripts/update-formula-checksums.sh
 
 # Cache snapshots / AOF written by a server started in the repo root
 # (e.g. while running scripts/benchmark.sh). Never artifacts to commit.

Formula/recached.rb

@@ -1,27 +1,33 @@
 class Recached < Formula
   desc "Blazing fast, multi-core drop-in replacement for Redis"
   homepage "https://github.com/recached-dev/recached"
-  version "0.1.8"
-  license "MIT"
+  version "0.2.4"
+  license "Apache-2.0"
 
+  # The checksums below are placeholders until the v0.2.4 release artifacts
+  # exist. Fill them with `scripts/update-formula-checksums.sh v0.2.4`, which
+  # downloads the published binaries and rewrites this file.
+  #
+  # `scripts/bump-version.sh` resets them to placeholders on every bump, and
+  # that is deliberate: this formula sat at 0.1.8 with *valid* 0.1.8 URLs and
+  # checksums while the project shipped 0.2.x, so `brew install recached`
+  # silently succeeded and handed people the old binary — including the one
+  # whose replication port served the keyspace without authentication. A
+  # placeholder makes brew fail loudly, which is the far better failure.
   on_macos do
     on_intel do
-      url "https://github.com/recached-dev/recached/releases/download/v0.1.8/recached-macos-x86_64"
-      # shasum -a 256 of the binary in target/dist/recached-macos-x86_64;
-      # recompute if the release binary is rebuilt before uploading.
-      sha256 "227fca7d7ff5c9511f9482863024e222c0c66f93ad40288581ca1fb32a9f20bd"
+      url "https://github.com/recached-dev/recached/releases/download/v0.2.4/recached-macos-amd64"
+      sha256 "REPLACE_WITH_AMD64_SHA256"
     end
     on_arm do
-      url "https://github.com/recached-dev/recached/releases/download/v0.1.8/recached-macos-arm64"
-      # TODO: build on Apple Silicon (or cross-compile: cargo build --release
-      # --target aarch64-apple-darwin), upload, then: shasum -a 256 recached-macos-arm64
+      url "https://github.com/recached-dev/recached/releases/download/v0.2.4/recached-macos-arm64"
       sha256 "REPLACE_WITH_ARM64_SHA256"
     end
   end
 
   def install
     # Rename the downloaded binary to 'recached-server' and install it into the Homebrew bin
-    binary = Hardware::CPU.arm? ? "recached-macos-arm64" : "recached-macos-x86_64"
+    binary = Hardware::CPU.arm? ? "recached-macos-arm64" : "recached-macos-amd64"
     bin.install binary => "recached-server"
   end
 

README.md

@@ -101,7 +101,7 @@ password, no TLS, and no restriction on which web pages may open the sync socket
 
 ## Benchmarks
 
-Measured with `redis-benchmark` (100k requests, 50 connections, 64-byte values, randomized keys, persistence disabled on all servers) on a 4-core Intel i5-8259U laptop, July 2026 — Recached v0.1.8 vs Redis 7.2.5 vs Valkey 9.1.0, one server at a time.
+Measured with `redis-benchmark` (100k requests, 50 connections, 64-byte values, randomized keys, persistence disabled on all servers) on a 4-core Intel i5-8259U laptop, July 2026 — Recached v0.1.8 vs Redis 7.2.5 vs Valkey 9.1.0, one server at a time. Current release is v0.2.4; these command paths were A/B tested across the v0.2.4 changes and moved within run-to-run noise, but the three-way suite has not been re-run since v0.1.8.
 
 Pipelined (`-P 16`) — raw command throughput, requests/sec, **bold** = best per row:
 
@@ -117,6 +117,8 @@ Pipelined (`-P 16`) — raw command throughput, requests/sec, **bold** = best pe
 
 Recached's multi-threaded runtime spreads connections across all cores, while Redis and Valkey execute commands on one — pipelined, Recached comes out ahead of Redis on 6 of 7 commands and ahead of Valkey on all 7, on the same hardware. Unpipelined (one command per round-trip — the traffic shape of typical request-scoped cache calls), the localhost round-trip dominates and Recached runs at 46–96% of Redis with sub-millisecond p50 latency on every common command (GET 58.1k vs 61.6k rps; HSET is the weakest at 46%).
 
+**New in v0.2.4:** sorted sets gained a score-ordered index, so range reads no longer sort the whole set on every query. On a ~45k-member leaderboard, repeated `ZRANGE key 0 9` went from 244 to 133k rps (**546×**), and an alternating `ZADD` + `ZRANGE` loop from 65 s to 0.11 s (**597×**). `ZADD` gives up ~15% against a set that is actively being read; a write-only sorted set never builds the index and is unaffected. Measured as before/after ratios on a loaded machine — see the [benchmarks page](https://recached.dev/guide/benchmarks) for methodology.
+
 Full tables with latency percentiles, pipelined results, methodology, and known hotspots: **[recached.dev/guide/benchmarks](https://recached.dev/guide/benchmarks)**. Reproduce with [`scripts/benchmark.sh`](scripts/benchmark.sh) — results from server-grade hardware welcome.
 
 ---