Darren's Daily Briefing

High-volume hotel price-cache architecture

Agoda moved a 1.5 TB price cache from 72 SQL Server shards to two DragonflyDB clusters and reports roughly eightfold lower P99 read latency.

The migration used dual reads, parity checks, staged traffic shifts, and decentralized failover detection to validate correctness before cutover.

Watch sustained tail latency, update correctness, failover behavior, operational cost, and whether the design holds at peak travel demand.

Windows kernel driver vulnerability research

DeepZero’s open-source pipeline narrows driver binaries through seven stages before asking a model to assess exploitability; the maintainer reports several verified findings.

PE parsing, IOCTL detection, known-driver filtering, headless Ghidra, Semgrep, and candidate ranking reserve model analysis for a smaller evidence-rich set.

Watch independent reproduction, disclosure outcomes, false positives, hardware-dependent reachability, and reuse of the pipeline on other binary targets.

Linux fleet migration and policy compatibility

Amazon Linux 2027 preview changes SELinux from permissive to enforcing by default, so applications that previously logged policy violations may fail.

Test file contexts, ports, access patterns, base images, and replacement-host workflows; AWS has not yet announced GA or an in-place migration path.

Watch application-policy failures, image availability, AWS-LC compatibility, rollout strategy, and support-lifecycle details before fleet migration.

Real-time analytics cost and performance benchmarking

ClickHouse’s own benchmark compares four cloud analytics systems on a stream of over 100 billion rows at a target one million rows per second; its reported advantage is vendor-authored.

CostBench includes ingest, query-readiness maintenance, and concurrent query serving; its score normalizes active compute and excludes storage, discounts, idle capacity, and fallback cost.

Watch independent replication, broader workloads, complete production bills, ingestion freshness, and whether result rankings persist under different query and retention patterns.

Production LLM inference and GPU energy economics

A September 10 preprint reports that one power profile across prefill and decode lanes yields only 8.6% more tokens per joule with 5.2% mean latency cost in a disaggregated B200 test.

The proposed controller sets an SM-clock window for compute-heavy prefill and a calibrated power cap for memory-bound decode, guarded by inter-token-latency targets. In one MoE test it reports 20.4% more tokens per joule at 3.5% mean latency cost.

Watch independent replication, tail-latency compliance, energy savings under real workloads, and whether gains transfer to dense models; the authors report much smaller benefits for a dense model.