Final PIR (held since 10-01; preliminary on 10-14). From ~20:30 UTC Sep 30 to 02:15 UTC Oct 1 (02:00–07:45 IST), management operations failed intermittently for ExpressRoute Gateway, Application Gateway, Firewall, Bastion, Virtual WAN and VPN Gateway, and some upgraded ExpressRoute gateways lost connectivity. A temporary restriction pausing OS updates expired, and a large backlog of gateway updates resumed almost at once, flooding a shared regional control plane with restarts and config requests. An earlier performance change had already made each config operation slower, and throttling/load protection didn't hold. ExpressRoute instances were marked healthy before their routing config was restored, so updates moved on to the redundant instance and neither could fetch config; after ~30 min of failed retries gateways fell back to locally stored last-known-good config. Pausing updates, shifting traffic and adding capacity only helped temporarily; rolling back the performance change fixed it. Follow-ups run to Jun 2027 (staged resumption, stricter health checks, control-plane scale-out, better load-shedding). Retrospective livestreams Oct 23, 00:00 and 12:00 IST.
🔥 What Changed
Normal-sized Saturday after two missed editions. The held Azure gateway PIR 7Q30-010 is final; Inngest shows Go context cancellation collapsing a PgBouncer pool; Go 1.27 now reuses HTTP/1 connections after early body close; Anthropic details Scion, its git replacement at 30k merges/day.
Holds: Kafka 4.4.0 artifacts are on downloads.apache.org (vote passed 10-13) but no announcement yet — waiting for the release blog. grpc-go v1.85 still unreleased (latest v1.84.0; 1.85 branch active 10-16). Go #81957 CL 847406 and alt CL 844405 unmerged, backports #82189/#82190 open; #82177 fix CL 848985 under review; #82026, #81759 unchanged; #81864 tip-only. No new Go point release.
Published 2026-10-15. Twice on 10-10, both PgBouncer poolers in front of Inngest's primary DB lost all connections and couldn't rebuild them. Short-deadline lookups cancelled queries still waiting in PgBouncer; the Go driver closed the client side, PgBouncer later ran the abandoned query and then destroyed that server connection instead of returning it. Poolers open new connections one at a time, so connections died faster than they were replaced, and thousands of queued clients hitting their own deadlines kept the loop going. It started at the top and bottom of the hour, when cron work roughly triples query rate. Raising pool size didn't help (connection setup, not the limit, was the bottleneck), and the fallback to the provider's poolers hit NAT gateways that silently dropped connections, hanging queries up to 16 min. Fix: a deadline now ends the caller's wait without cancelling the query. Corroborated by pgx issue #1764 and open PR #2425.
context deadlines on Postgres calls behind a pooler are everyday Go backend code, and this shows cancellation alone taking down the database tier.
Published 2026-10-14 (Nicholas Husin, Damien Neil). Through Go 1.26, closing an HTTP/1 response body before fully reading it made net/http discard the connection — even when json.Decoder left only a trailing newline — so the next request paid fresh TCP+TLS setup. In Go 1.27, http.Transport drains an unread HTTP/1 body in the background for up to 50 ms or 256 KB, skipping it when the response says Connection: close. Their benchmark: 100 GETs to index.golang.org went from ~710 ms (1.26) to ~150 ms (1.27), about 5×. Recommendation is now just defer resp.Body.Close() and dropping hand-written drain helpers. Draining was never needed for HTTP/2, and the limits may change in future versions.
Published 2026-10-13 (updated 10-15). Monorepo merges grew from ~2,500/day in May to 30,000+/day in September; ~100 GB of history and ~25M checkouts/day (600+ PB/day if fetched naively) outgrew several hundred git mirrors, so a team of 7 built Scion and migrated off git in five months. Each commit is a write-ahead-log entry listing changed files with content-addressed blob pointers; a branch/tag update is a compare-and-swap on a single Spanner row, separate from commit creation. Blame, history and graph indexes are derived data, computed from parents eagerly or on demand. Checkouts flow per-node daemon → regional cache → authoritative servers, all answering the same RPCs with strictly enforced tiering for failure isolation and >99% hit rate per tier. Reported: 200× less CPU/memory than the mirror fleet, checkouts in seconds, 25% less CI CPU. Caveats: internal unaudited numbers, not offered externally, builds on Piper and Sapling/Mononoke ideas.
📚 Worth Your Time
Released 2026-10-07; v2.0.1 bug-fix on 10-14. Go API major only: Connect wire protocol and gRPC compatibility unchanged, and v1 lives at a different module path so both can coexist (v1 keeps getting fixes). Generic connect.Request[T]/Response[T] wrappers are gone; handlers look like gRPC-Go, func(ctx, *Req) (*Resp, error). The core no longer depends on net/http: connecthttp serves Connect, gRPC and gRPC-Web, and connectinprocess lets tests call a server without a network. One interceptor model covers unary and streaming and runs before decompression and decoding, so auth can reject on headers alone. Only errors created as *connect.Error send a message to clients, so DB error text no longer leaks. Core exports ~25 identifiers instead of ~180; connect-go-v2-migrate does most of the rewrite, but custom interceptors, otelconnect and vanguard need manual changes.