ClickHouse eng (Sep 21) — when Postgres working set exceeds RAM, storage latency masquerades as DB problems (ingest, p95/p99, VACUUM, checkpoints, logical-repl lag). Benchmark: identical m6id.4xlarge / Postgres 18.3 / shared_buffers=16GB / pgbench scale 33k (3.3B rows, 482 GiB, ~30× RAM): local NVMe vs gp3 EBS @ 3k IOPS. Result ~9.2× TPS (16,030 vs 1,734) and 4.0 ms vs 36.9 ms UPDATE latency. EBS left 45% of backends in IO:DataFileRead (~89% wall off-CPU); NVMe 14%. Durability: AZ-local ephemeral NVMe + quorum sync (synchronous_standby_names = 'ANY 1') + WAL-G to S3. Skim CDC / WalShadow / Managed Postgres product wrap.

Michael Stapelberg’s production write-up of deleting the last cgo dependency in Debian Code Search by reimplementing TurboPFor in pure Go with GOEXPERIMENT=simd / simd/archsimd. Covers zero-alloc streaming encoders, GOAMD64=v3/v4, PGO pitfalls, generics-as-bit-width specialization, AVX2/AVX512 kernels, and a positional popcount scan that ~2×’d the encoder after matching cgo. Still ~1.4× behind clang C; beats the prior cgo path DCS shipped. Documents remaining Go codegen gaps.