Daily Engineering Intelligence

Technology Newspaper

← Back to Home

Engineering Intelligence — Monday

November 2, 2026

🔥 What Changed

A quiet Monday with two items, both long reads. No Go, Kafka, Kubernetes, NATS or auth-library release or advisory worth a slot landed over the weekend; the GitHub Advisory Database has nothing new since early 10-31 IST. Both reads are about the same shift: LLMs are producing a lot of buggy code, and they are also finding bugs faster than maintainers can fix them. Jepsen's ParticleDB report is the more useful of the two.

Also seen: Go CL 857381 (external contributor, 11-01, under review) fixes an http.Request.Write framing gap. If you set TransferEncoding to anything other than chunked (for example identity, to avoid chunked uploads) and the body length is unknown, Go sends the body with neither Content-Length nor chunked framing. An HTTP/1.1 server then treats the request as having no body and reads the body bytes as the next request on the connection. The CL probes the body and returns an error instead. HTTP/2 is unaffected. If your Go clients set TransferEncoding by hand, set ContentLength too. Go #82717 (MPTCP on by default) has had only side comments since Saturday; no decision. No new Go, Kafka, NATS, Kubernetes or OCR release cleared the bar.

Holds: Go 1.27.4/1.26.11 (still not pre-announced; go.dev/dl lists 1.27.3/1.26.10); Go #82717 MPTCP default (no decision); CVE-2026-98432 RHEL/Ubuntu kernels; x/oauth2 CL 854305; Go CL 857381 (#82756); Go #82374, #82472, #82497, #82547, #82677, #82682/#82685 unchanged; go-jose v3 fixes; Kafka 3.9.3; GitHub RCAs for 10-21/22/26 and 10-28 to 10-30; Northflank root cause; Dapr CVE/GHSA IDs; Postgres 19 release date. Alternates: Pivot "It's the memory latency, stupid" (GROUP BY cache misses), AWS File Mover Express (InfoQ), AWS capacity largely sold out through 2027 (Jassy via InfoQ), Simon Willison on deleting uv.lock.

No release or advisory cleared the bar today.

📚 Worth Your Time

Jepsen tested ParticleDB, a Postgres-compatible database built largely with LLMs, and found 96 bugs, most of them on a single healthy node with one client

Kyle Kingsbury tested ParticleDB, the "multi-model database for AI workloads" from Orkes (the company behind Conductor), versions 1.1.4 through 1.1.18.11.rc-22. Orkes paid for the work. He ran single Debian 13 nodes in containers and on EC2, killed processes, simulated lost un-fsynced writes with LazyFS, and truncated or corrupted files. He found 96 issues, and all but three happened on a healthy single-node instance, often with one client and no faults at all. The list covers every transactional anomaly the checker can detect (write loss, G0, G1a/b/c, G-single, Lost Update, G2 and real-time variants), plus SQL that was simply wrong: WHERE 4 = 5 returned the row with primary key 5, NOT NULL and UNIQUE constraints did nothing, primary keys could be duplicated or NULL, aborted transactions could commit, and rows from dropped tables came back in mangled form. Corrupting the WAL lost some data; simulated misdirected reads and writes could lose all of it. By rc-22 all but ten issues looked fixed, but the default occ mode still lost committed writes in normal operation, and serializable still allowed stale reads, plus Lost Update after a crash. Orkes says 1.1.19-rc.13 fixes everything; Jepsen has not tested it. Most of the bugs came from about 2,000 lines of generative tests that run random schemas and SQL against both ParticleDB and real Postgres and compare the results. Jepsen is now reusing that SQL generator in other tests. The Discussion section is the part to read. Orkes said it had already run Jepsen's Postgres harness and SQLancer, but that harness failed immediately against 1.1.4. The docs claimed a "formally verified" TLA+ proof that was never published, and they invented Java package names, default usernames and an email address for Kingsbury. In a section titled "LLM Tradeoffs", Kingsbury writes that LLM tooling repeatedly claimed bugs were fixed when they were not, and that Orkes could not explain the cause or the fix for nearly any issue: new builds arrived without changelogs. Jepsen's advice to users: stop deploying ParticleDB, migrate, and rebuild state from a more reliable source.

Why you care: You won't run ParticleDB, but you will review code and evaluate infrastructure that an LLM wrote much of. This is the most concrete public evidence so far of how that goes wrong. A polished website, a TLA+ claim, thirteen client libraries and "Full ACID" all turned out to be cheap to fake, while simple example tests (insert a NULL into a NOT NULL column and check it fails) were missing. Two habits carry straight over to Go services. First, differential testing against a reference implementation (your new store against Postgres, a new parser against the old one, a rewritten consumer against the existing one) found subtle bugs with a small amount of code. Second, "the model says it's fixed" is not evidence: ask for a failing test that passes after the fix. If you're choosing a database or queue for an event-driven system, look first for an independent Jepsen-style report and a real changelog.
Read Jepsen: ParticleDB 1.1.4 →

Robert Haas: Postgres 19 is late because LLMs are finding old bugs faster than committers can fix them

PostgreSQL major releases have reliably shipped between mid-September and mid-October for about a decade. Long-time committer Robert Haas says Postgres 19 most likely won't ship until November, and the project could postpone it further. Three large features were also pulled late in beta, between August 27 and September 15: UPDATE/DELETE FOR PORTION OF, SQL/PGQ property graph queries, and ALTER TABLE … MERGE/SPLIT PARTITION. His numbers, comparing April 8 to September 25 each year: commits back-patched to the release branch rose from 724 (REL_18_STABLE, 2025) to 1,169 (REL_19_STABLE, 2026), and commits back-patched to the previous stable branch as well rose from 248 to 433, up more than 74%. New work on master stayed about flat. His explanation: many people now report bugs in Postgres that they found with LLMs. LLMs can draft fixes, but humans still have to review and commit them, and that review time came out of preparing v19.

Why you care: Read this next to the Jepsen report. LLMs are finding bugs in mature, carefully reviewed code faster than humans can fix them, while also producing new code with many bugs. Expect the same in your stack. Go's queued batch of 12 private stdlib CVEs and this month's long run of Kafka and Go library advisories may be part of the same trend. For planning: allow for more frequent patch releases from your upstream dependencies, keep your dependency-bump pipeline (tests, staged rollout) cheap enough to run weekly, and don't count on Postgres 19 features landing on the usual date. If you maintain a Go library, expect LLM-found reports and decide now how you'll triage them.
Read Robert Haas's post →