P99 CONF 2026: Performance Engineering Opens With Two Keynotes About Agents
Table of Contents
Event
| Field | Value |
|---|---|
| Event | P99 CONF 2026, the sixth edition |
| Organiser | ScyllaDB — "the organization that curates P99 CONF" |
| Dates | Wednesday 21 — Thursday 22 October 2026 |
| Hours | 08:00–13:00 Pacific both days, plus a 01:00–04:00 Europe/Asia/India track on day two |
| Format | fully virtual, three concurrent stages; talks recorded in advance, Q&A live |
| Cost | free; registration carries 30 days of the O'Reilly platform |
| Scale | 62 talks, 14 of them still unscheduled; "thousands of engineers attend each year" |
| Registration | lp.scylladb.com — no deadline stated |
Its own framing, from the homepage:
There's no other event like this -- a conference for engineers by engineers, where we'll share novel approaches for solving complex problems efficiently and at speed. Vendor and tool agnostic, this conference will be for a highly technical audience only. Your boss's boss is not invited.
The CFP adds the constraint that makes the vendor-neutrality claim testable: "It's vendor-neutral, and we don't allow shallow overviews or product pitches." Talks are 15–20 minutes.
Why this one matters here
A conference about tail latency has put coding with agents in both day-one keynote slots and the day-two opener. That is the finding, and it is about where the subject has landed rather than about any single talk.
- Effective Coding and Working with Agents — Chip Huyen, author of AI Engineering. On making "the most out of working with multiple agents".
- Hardwood: Building a Parquet Parser From Scratch (With a Little Help From AI) — Gunnar Morling, Confluent. The abstract names the tool: "practical learnings from using AI (specifically, Claude Code) as a coding companion".
- How to Improve Your Cache Algorithm Using AI — Dor Laor, ScyllaDB's CEO, opening day two. "Everybody knows that AI can code. It can also convert your code to Rust. But can AI write a database? … my journey of 'vibe coding' the ScyllaDB cache to implement an alternative LRU algorithm and improve our p99 latency."
Roughly eighteen of sixty-two talks touch AI. Two are close enough to the work here to be worth the timeslot.
The agent-in-production talks
The Autonomous Performance Agent: A Netflix Production Story — Rajat Shah, AI Platform at Netflix. An agent that "continuously hunts performance inefficiencies across live production services, traces them to source code, proposes fixes, and validates results through canary deployment — grounding every decision in measured production outcomes, not model confidence". The abstract's own framing of the hard part: "what it actually took to make an autonomous agent trustworthy enough to act in production: where it earns autonomy, where it doesn't".
That is a calibration question wearing production clothes. An agent that proposes a fix and validates it against a canary has an oracle. The interesting claim is where the boundary sits.
Give the Agent a Cluster: Effective AI for Performance Engineering at Scale — Filipe Oliveira, Redis. Currently unscheduled. "What happens when you give an AI agent real power: access to performance profiles via MCP servers, a full multi-core cluster for controlled experiments, and a dedicated budget for iterative code generation and optimization. Instead of asking, 'Can AI write code?', this talk explores a more interesting question: Can AI run your performance lab?"
Profiles over MCP is the same move as exposing a REPL over MCP, applied to a different instrument. Worth reading against the local notes on expanding what an agent can see of a running system.
And one number to check
Vectorless RAG: How Removing the Vector Database Cut P99 by 40% — Jubin Soni, Yahoo. Unscheduled. "For a meaningful class of RAG applications, the vector database is the slowest part of the pipeline… In some production systems, we've seen, 40% P99 reductions after removing vector search, with recall improving rather than degrading."
A claim that removing the component improves both latency and recall is the kind that deserves its methodology read before it is repeated.
The archive is the durable part
p99conf.io/on-demand carries 264 session pages spanning 2021 through 2025, filterable by year, type and topic, each with abstract, speaker bio and slides. Ungated — an attendee quote the site chooses to display reads "Videos available on-demand afterwards with no gating or games."
For a conference whose talks are recorded in advance anyway, the archive rather than the live event is the artifact. ScyllaDB also runs a separate gated on-demand landing page for 2025, which is worth knowing before handing over an email.
Discrepancies on the published material
Recorded because a reader planning around this page would hit them.
- The UTC conversion is wrong. The site says "8:00am – 1:00pm Pacific Time
/ 16:00 – 20:00 UTC". October Pacific is UTC-7, so 08:00 PT is 15:00 UTC,
not 16:00. Sessionize's own timezone field says
UTC-07. - The stated hours exclude a whole track. "8:00am – 1:00pm" on both days, while the grid schedules nine day-two sessions from 01:00 to 03:55 Pacific for Europe, Asia and India. Day two actually runs about 01:00 to 12:35.
- "Full agenda will be announced soon" still sits on the homepage while
/agenda/renders a complete two-day, three-stage grid. - Three talks carry different titles on the homepage and in the agenda, including one where the vector count differs — 1B on the homepage, 100M in the abstract.
- The lower speaker wall is a legacy reel. Andy Pavlo, Michael Stonebraker, Gil Tene, Liz Rice, Charity Majors and others appear with bios and no 2026 talk. They are past speakers used as social proof and read as this year's lineup.
- One talk exists only on the homepage — Catherine Jue of KERNEL on "building crazy fast, open source infrastructure for AI agents", with sub-30ms browser boot times — and appears nowhere in the agenda feed.
Sources
- p99conf.io — server-rendered; the homepage carries the speaker list and abstracts
- CFP — closed 29 May 2026; carries the track list verbatim
- On-demand archive, 2021–2025
- 2025 recap — "the fifth in this incredible series", titled Latency to LLMs