The Raft consensus log, dissected at the byte level.
A line-by-line reading of the canonical Raft implementation, the seven places it can lose data, and the test harness we used to prove it.
Depth over hype — written by working seniors, peer-reviewed by a second, benchmarked on reproducible hardware. Shipped every Tuesday since 2012.
Editor's note
The internet does not need another algorithm-grind interview blog, and it certainly does not need another AI-generated rewrite farm. TechProgramming is a working engineer's alternative — a quiet, slow publication run by a 42-person editorial collective of senior engineers from Google, Stripe, Mozilla, Jane Street, and Cloudflare.
Every article is written by a practicing senior engineer, peer-reviewed by a second, and benchmarked on reproducible hardware in our in-house runtime lab (ARM, x86, WASM). The median piece runs 4,200 words and 11 minutes 40 seconds of reading. Nothing ships until two humans have argued about it and the data is published with the article — not behind a paywall, not in a footnote, not summarised into a marketing takeaway.
David Reimer & Priya Chandrasekar, founders · Berlin, 2024
This week's issue
Six long reads, four reference sheets, one benchmark drop. Shipped Tuesday, peer-reviewed by Thursday, on the page by Friday morning.
From the TLB miss to the swap-in path, traced on kernel 6.8 with perf, bpftrace, and a deliberately misbehaving workload.
DMB, DSB, ISB, the load-load, load-store, store-store quadrants, and why your CI passes on x86 but flakes on Graviton.
Key update, 0-RTT replay, early data rejection, and the four ways a session resumption silently breaks production.
A 48-hour benchmark with raw IO traces, the configuration knobs that actually moved the needle, and the ones that did nothing.
From the archive
The full index averages 4,200 words per article. The four below are the ones readers return to most often.
A line-by-line reading of the canonical Raft implementation, the seven places it can lose data, and the test harness we used to prove it.
We benchmarked seven libc implementations across three ISAs and published the raw traces. The result is not the one the benchmarks usually claim.
A working engineer's reading of AES-GCM, ChaCha20-Poly1305, and the four foot-guns that appear in production code review every quarter.
How a misconfigured kernel accounting module masqueraded as an application leak across three on-call rotations, and the diagnostic that finally cracked it.
From a recent letter to the editor
“Most engineering publications are written for people who want to sound senior in a meeting. TechProgramming is written for people who actually have to ship the thing at 3am. I cancelled four other subscriptions the week I subscribed.”
— A staff engineer at a major infrastructure company, in correspondence shared with the editor's permission
The publication, in metrics
One Tuesday email per week. No tracking pixels, no “suggested for you”, no affiliate rewrites.