Final PIR on Azure status history (incident 2026-09-29 Sweden Central; ~10:03β15:58 UTC). A planned internal-service deploy cleared caches at peak β cache-miss defects amplified lookups β backend traffic ~20Γ normal β caching layer saturated β outbound connection capacity exhausted. Automated instance health checks required DB connectivity; failed checks restarted instances β fewer healthy replicas β retry amplification. Mitigation: reconfigure health check to stop depending on remote DB connectivity (completed). Follow-ups: throttle/rate-limit backend, grow/partition cache, better connection-exhaustion diagnostics. AIR livestreams Oct 22/23.
π₯ What Changed
Normal Wednesday. The held Azure OpenAI/Foundry PIR 7QL5-Z50 lands as a final write-up (cache stampede β connection exhaustion β health-check restart loop). Cloudflareβs D1 control-plane post shows how to keep millions of per-tenant SQLite DBs off a central Postgres hot path. Selective OCR: Datalab ships Indra (~1M-param ONNX, ~49 pp/s on 32 CPU cores).
Holds: Azure 7Q30-010 still Mitigated only (PIR not yet; expect ~10-15). Go #82026 CL 846325 still NEW unmerged; #81957 Pratt CL 847406 still NEW (backport issues #82189/#82190 opened 10-13 β story promoted to WYT today, merge not yet); #81759 open CL 840845; #81864 tip-only, no backport. grpc-go v1.85 still not released (milestone was 10-13; latest tag v1.84.0). #82177 open; #82120 closed not_planned; #82146 open (incremental β not resurfaced).
Published 2026-10-13. D1 = per-DB SQLite Durable Object; 15Γ live DBs YoY. Guiding rule: no centralized single-region component on the query hot path. Workers KV holds DBβDO routing (durable, not just cache); Hyperdrive+Postgres for control-plane CRUD; Edge API β Eyeball β DO for queries. Creation: init DO then Postgres checkpoint; heartbeats (~30s) self-heal orphaned DOs and KV. Usage limits: DO β queue β per-DB stats β account pre-aggregation (avoid SUM over millions of rows at request time). Binding path skips UUID lookup.
Published 2026-10-13 (blog + GitHub + PyPI 0.3.1). CPU-first package: line boxes, layout-block labels, reading order, markdown/JSON; digital PDFs use text layer when usable. Vendor OmniParseBench headline 62.1 vs PP-StructureV3 58.4 at ~1/180 the time; 49.4 pp/s on 32 cores. Opt-in --math. Caveats: code Apache-2.0; weights modified OpenRAIL-M (commercial bar if org revenue/funding >$20M); vendor bench; weak complex tables/handwriting; scan OCR scripts limited.
π Worth Your Time
Filed 2026-10-02; open; assignee prattmic. HTTP reverse proxy under spike: thousands blocked on one sync.Mutex; runtime.adjustsudogs ~43% of CPU via shrinkstackβcopystack. Regression from goroutine-leak gp.waiting = s on sema queues (waitlist = other waiters, not channels). Repro: 40k waiters GC CPU ~0.75s (1.25) β ~9.8s (1.26/1.27); GODEBUG=gcshrinkstackoff=1 restores. CL 847406 (only adjust sudogs for channels) still NEW; Pratt requested backport to 1.26/1.27 on 10-13 (#82189/#82190).