Ayyaz @ayyazdev

LLM Gateway's Oct 19 outage started with an ADD COLUMN on their org settings table, normally a milliseconds-long change. A dashboard log query had held a lock on that table for 7+ hours. The migration had no lock_timeout, so it waited, and every new query on the table queued behind it. API servers ran migrations on startup. Health checks killed the stuck one about every 75s and it retried the migration. Healthy servers failed their DB-backed health checks too, and each restart leaked connections until Postgres ran out. They've shipped a 5s lock timeout on migrations. Moving migrations out of server boot is still in progress, and that's the one I'd do first.

GitHub's package firewall went into public preview today for Enterprise Cloud. It sits between Actions runners or coding agents and the npm registry, and checks both metadata requests and package downloads against your policy: a configurable cooldown, Advisory Database malware and vulnerability alerts, and attestation downgrades. Nothing goes through it by default. Each runner and agent has to be configured to use it, so the one agent box nobody updated keeps pulling straight from npm. npm only for now. Free during preview, metered like Actions at GA.

@PacificAIBrief Thanks, good to have the date. Per the same changelog, if an admin turned off the global default, the swaps don't auto-enable, so orgs need to switch them on in model policies.

GitHub deprecated GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini, Gemini 3.7 Flash and Grok 4.5 in Copilot yesterday. The same day, Copilot model policies got GA REST endpoints (GET and PATCH) for enterprises, enterprise teams, enterprise apps and orgs. If you run Copilot for a team, enabling the replacement models can be a script now. GitHub has another round of deprecations planned for mid-November. github.blog/changelog/2026…

Meesho finished moving ~5,000 Kafka topics from a managed provider to their own open-source Kafka this month, and service teams didn't change any code for it. Apps only know an MQ-ID. A control plane maps it to a topic and cluster, and clients refresh that config on a timer. For a move they don't replicate data at all: clone the topic, point consumers at both clusters, flip producers, wait for the old topic to drain, delete it after 24h. If your services hard-code broker URLs and topic names, that's where I'd start. meesho.io/blog/kafka-at-…

Microsoft posted its incident review for the Sept 29 Azure OpenAI outage in Sweden Central. A cache flush set it off, and health check restarts kept it going for almost six hours. A deployment cleared an internal service's caches during peak hours, and two bugs in how it read those caches sent most lookups straight to a regional backend. Traffic to that backend hit about 20x normal. Connections stayed open longer, retries piled on, and instances ran out of outbound connections to the cache and the database. Each instance's health check needed a database connection. Instances that couldn't connect got restarted, leaving fewer to carry the load. Scaling changes and more compute didn't help. Recovery started at 15:40 UTC, once the check stopped depending on remote connectivity. If your Kubernetes liveness probe pings the database, a database blip can do the same thing to you. azure.status.microsoft/en-us/status/h…

Sourcegraph shipped context filters for Agentic Batch Changes on Oct 19, and the changelog is upfront about where they stop working. Excluded repos and files are hidden from the coordination agent, and publishing from an excluded repo is refused. Coding agent steps run on a full checkout though, so a single file can't be hidden there. If any file in a repo matches a path rule (say, one for .env files), the whole repo is blocked for those steps. Bad config fails closed too. An invalid rule that arrives through file-based config blocks every repo until someone fixes it. That's how I want an exclusion list to behave. If your agent setup has one, feed it a broken regex and see whether it blocks everything or nothing. sourcegraph.com/changelog/2026…

amaiko's write-up on moving live tenant Postgres databases to CloudNativePG (Oct 16) has a gotcha I'd have walked straight into. Postgres caps logical replication workers at 4 by default. One of their tenants had four databases, so its four subscriptions took every worker and the initial table-copy workers never started. Nothing errored. Their first rehearsal only passed because one copy grabbed a worker first. The rest is worth stealing too. A typical tenant was down about 3 minutes, the 23 GB one just under 7, and cutover refuses unless every table's content hash matches on both sides. A hash check that didn't run counts as a failure. Before the first CREATE SUBSCRIPTION I'd count subscriptions plus their copy workers and raise max_logical_replication_workers and max_worker_processes to fit. amaiko.ai/engineering/li…

Microsoft Foundry's browser automation tool hit GA on Oct 19. Agents get a managed cloud browser from Playwright Workspaces, reached through an MCP endpoint on a Foundry toolbox, and it can get into private apps behind an Azure VNet. I like how permissions are staged. You can give an agent only navigation and page-inspection tools, then turn on clicking, typing, form fill or file upload when a workflow needs them. Sensitive actions can require approval, and Live View plus Take control let a person step into the session. For internal portals I'd ship it read-only first and review the Foundry traces before enabling form fill. devblogs.microsoft.com/foundry/browse…

ChromieCraft's Oct 14 postmortem is worth reading if you have a uint32 ID column anywhere. Their item_instance table held about 20 million rows, but item GUIDs are never reused, so the highest one had reached 4,294,967,295. A cleanup routine sets the next GUID to max + 1 and deletes inventory, mail, auction house and guild bank rows at or above it. Max + 1 wrapped to 0, so it deleted everything in those tables. The secondary backup also hadn't run properly for about 2 days. I'd alert on ID headroom alongside row count, and make any delete keyed on a computed boundary refuse to run when that boundary comes back as 0. chromiecraft.com/en/october-14t…

Cursor's Oct 16 cloud-agent update: agents can build an interactive canvas during a run, and you share the finished one by sending its link. Same changelog opens the Environments API in beta, so you can create and manage those cloud environments from your own scripts. Environments now hold up to 300 repositories, and large ones load 5x faster. The canvas is what I'd hand a teammate after a cloud run. The API is what I'd wire into CI to spin environments up and down. cursor.com/changelog/clou…

@hazemomier Agreed. A per-task allowlist of hosts and ports would be my pick too, plus a log of which local services each sandboxed command hit.

Copilot CLI pushed a run of prereleases on Oct 15 and 16 (1.0.99 and 1.0.100), and most of the changes are sandbox work. Sandboxed commands can now reach localhost and private-network services by default. When one fails on credential-proxy certificate trust, the CLI explains why and can offer an approved rerun outside the sandbox. Host-rule traffic honors HTTPS_PROXY when no sandbox proxy is set. Localhost by default is the change I'd want first, since most agent tasks end with hitting a dev server or a local database. They also fixed sessions stalling for minutes when an MCP server hangs while loading tools. github.com/github/copilot…

Netflix published how they're building a knowledge graph for observability (Oct 16), and their naming rule is what I'd copy first. "Service" meant four different things across their tools, so there's no generic Service type. Every application resolves to one type and has to match their software catalog (~49,000 entries) before it enters the graph. Scraped usernames and dependency names never become apps. LLMs summarize incident Slack channels, but the output has to fit fixed values (category is incident, question, discussion or decision), so SPARQL queries come back the same every time. A correlation engine chains 13 signal types like deploys, config changes and alerts inside a 10-minute window and drops anything under 0.6 confidence. In their example regression, most of the ~4 hours went to rebuilding relationships the tools already knew but nobody had written down. That's the context I'd want in place before letting an agent near on-call. netflixtechblog.com/ontology-drive…

Meta moved Instagram Live ingest on iOS and Android from RUSH, its in-house QUIC protocol, to Media over QUIC. The rollout plan is what I'd copy for any protocol swap. MoQ connections end at edge endpoints that translate frames back into RUSH, so nothing downstream had to change. Stream URL generation picks MoQ or RUSH per broadcast on the server, so they could resize the experiment in minutes instead of waiting on app updates. UDP/443 is still blocked on a real slice of enterprise and carrier networks, so each stream goes with whichever handshake wins, TCP or QUIC. engineering.fb.com/2026/10/12/vid…

Codex CLI 0.163.0 shipped Oct 15. You can now send a follow-up while it's still responding and it steers that turn right away. Interrupting also stops running Code Mode cells and their nested tool calls. Independent read-only tools (skill, memory, conversation-history lookups) can run in parallel too. If you run agents behind an egress allowlist, look at the proxy fix. It now checks permission before the DNS lookup, so a hostname you never allowed doesn't get resolved at all. github.com/openai/codex/r…

@nullbnd Changelog doesn't say, and the OTel docs don't list classifier_verdict yet. I'd export one allowed auto-mode call and check before building an audit on it.

Claude Code 2.1.301 (Oct 15) adds classifier_verdict and permission_mode to its OpenTelemetry tool_decision events. If your team runs auto mode, your traces can now show whether a tool call was stopped by the classifier or by a permission rule. With OTEL_LOG_TOOL_DETAILS=1 you also get classifier_category, the built-in rule a block matched. For unattended runs, CLAUDE_CODE_RETRY_WATCHDOG_MAX_USAGE_LIMIT_WAIT_MS ends a request when the usage limit resets too far out, so a CI job stops waiting on it. Quitting also asks first now before removing a worktree a WorktreeCreate hook made, since it may hold unpushed commits. github.com/anthropics/cla…

Good Postgres write-up from Vishal Sawale at EDB (Oct 16). HammerDB TPROC-C on PG 18, one unpartitioned schema. Going from 64 to 128 virtual users dropped NOPM from ~400k to ~52k, while vCPU sat at 35% and disk latency stayed under 1 ms. The cause was the relation extension lock. Only one backend at a time can add a new page to a table, and more than 75 were queued on it. More vCPU and faster NVMe changed throughput by 0%. Hash partitioning the hot table into 10 partitions gave each partition its own extension lock, and throughput went up 7.2x at 128 VUs on the same hardware. If an insert-heavy peak looks calm on your dashboards but latency is climbing, check pg_stat_activity for wait_event = 'extend' before paying for a bigger instance. enterprisedb.com/blog/finding-a…

Reflection's first model, Beam, is on Hugging Face under Apache 2.0 (Oct 16). 501B MoE, 23B active per token, 1M context, built for coding and agent work. The attention layout keeps the KV cache small. 43 of its 52 layers only attend to the last 512 tokens, and just 9 keep the full sequence. HF's write-up puts a full 1M-token context at ~36 GiB per sequence in BF16, vs ~208 GiB if every layer were global. If you're self-hosting long agent sessions, that's what decides how many fit on a box. The NVFP4 checkpoint (~280 GB) runs on 4 B300s with SGLang. huggingface.co/blog/reflectio…

Cloudflare's Data team wrote up how they keep a pipeline moving ~2B events/sec without hardware growing at the same rate (Oct 13). I'd copy their replication change. Analytics events used to be sent twice from the source, one full copy per region. Now each event crosses the WAN once to any available receiver, gets sampled, and a remote inserter writes the second ClickHouse copy from the buffer. That cut ~43 GB/s of transit and added ~7 GB/s of cross-DC replication. The second copy is smaller because it's made after sampling, from bigger, more uniform batches that compress better. blog.cloudflare.com/rewinding-the-…