PlanetScale’s TIN (“Text INdex”) is a GA Postgres full-text extension that uses native heap ctids as postings instead of contiguous segment-local doc IDs. Design spine: two-level page/offset bitmaps for discontinuous 48-bit ctids; AVX2/AVX-512 work elision (AND page bitmaps before decoding offsets); MVCC via heap checks + visibility-map intersection + VACUUM-cleared liveness bitmaps; segment merge without renumbering. On an 85 GB / 150M-doc Stack Exchange corpus vs ParadeDB / pg_textsearch / GIN: mixed top-10 199 vs 7.9 QPS, p99 256ms vs 6.7s, ~65 vs 582 MB/query.

Major Cilium release (CNCF, Sep 14). Platform headlines: Gateway API v1.4→v1.6 with ExternalAuth, TCPRoute/UDPRoute for L4 on the same Gateway as HTTP, ListenerSets, CORS filters; datapath plugins (beta) to extend eBPF without forking; ENI IPAM IPv6; Maglev EndpointSlice weights + weight-0 drain; ztunnel sidecarless mTLS ops; ClusterNetworkPolicy Admin/Baseline tiers; MCS-API stable for ClusterMesh (ServiceExport / clusterset.local); cluster-pool→multi-pool IPAM in-place migration; cilium-cni binary −79%.

Lorin Hochstein’s resilience analysis (Sep 19) of GitHub’s 2026-09-13 shared-auth DB outage. An offline data-cleanup job wrote to a primary serving nearly every authenticated request; its safeguard watched only replica lag and missed primary max_connections saturation; long DB timeouts let blocked handlers accumulate; token-creation retries kept the primary pinned. Forward fixes: rate-limit background jobs on customer-serving DBs, page on primary load not lag alone, bound token retries, request-level timeouts, break apart the shared auth cluster.

Figma replaced PgBouncer with PGKeeper, a Go service exposing gRPC to DBProxy. Three stages: admission control → connection pool (warming + token-bucket rate limits on create and teardown) → query execution. Load management via bradenaw/backpressure: priority semaphore + CoDel/adaptive LIFO, then a V2 multi-dimensional fair-sharing tree. Connection hygiene: bounded exhaust, auto-ROLLBACK, context→pg_cancel_backend via a dedicated cancel pool. Claims: DB SLO >99.99%; >20 user-visible incidents prevented in Q4 2025 alone.