Blog
Writing on GPUs, LLMs, MLOps, Kubernetes — and mindset · 3525 posts
#2026-03 765#english 592#culture 264#deep-dive 254#kubernetes 249#career 229#ai 219#llm 209#devops 196#2026-04 146#security 143#database 114#observability 113#communication 109#history 107#architecture 100#productivity 96#finance 88#economy 84#mindset 81#psychology 80#ai-papers 79#food 78#it 78#travel 78#deep-learning 77#japanese 77#networking 77#performance 72#business-travel 70#linux 70#gpu 69#ai-agent 66#cs-fundamentals 63#postgresql 61#rag 58#self-improvement 55#learning 53#mlops 53#python 52
OTel's Kubernetes Attributes Are Stable Now — What to Do Before k8sattributes Flips Its Default
OpenTelemetry's Kubernetes semantic conventions were promoted to stable in semconv v1.42.0 on June 12, 2026. But the Collector's k8sattributes processor still defaults to the old schema (v0), so most people haven't notic
2026-07-16 · 12 min read #opentelemetry#observability#kubernetes#semantic-conventions#telemetry-pipelineNative Histograms Are Stable Now — So Why Haven't You Turned Them On Yet
Prometheus native histograms showed up as an experimental feature in v2.40 (November 2022), went stable in 3.8.0 (December 2025), and 3.13, released July 1, 2026, is the first LTS to ship the feature in stable state. The
2026-07-16 · 17 min read #prometheus#observability#native-histograms#metrics#monitoringAWS Lambda MicroVMs — Running Agent-Written Code Under VM Isolation, and the New Problem Snapshots Create
AWS launched Lambda MicroVMs on June 22, 2026. It sells Firecracker micro-VMs as a service, and the market it's aiming at is clear — running user- or AI-generated code, isolated per tenant. It moves the container's share
2026-07-16 · 17 min read #security#firecracker#microvm#aws-lambda#sandboxingMCP Drops Sessions — Reading the Stateless Core in the 2026-07-28 Revision
The next MCP specification revision, 2026-07-28, is the biggest change since launch. Its core move is stripping state out of the protocol layer — the initialize handshake and Mcp-Session-Id sessions disappear, and every
2026-07-16 · 16 min read #mcp#ai#protocol#agents#integrationpg_plan_advice in Postgres 19: The Compromise From a Project That Has Long Refused Hints
The PostgreSQL project has refused optimizer hints for a long time. But PostgreSQL 19 Beta 1, released on June 4, 2026, added two contrib modules built by Robert Haas: pgplanadvice and pgstashadvice. This isn't Oracle-st
2026-07-16 · 14 min read #postgresql#query-planner#query-optimization#database#performanceParquet 2.13.0 Adds nan_count and IEEE 754 Total Order — How Float Statistics Got Fixed After 3 Years and 3 Months
Parquet's min/max statistics are the mechanism that lets query engines skip data. But on floating-point columns, this mechanism has been subtly broken for over 10 years — NaN returns false in every comparison, so it gets
2026-07-16 · 17 min read #data-engineering#parquet#columnar-storage#apache-arrowDo You Need a New Learning Rate When You Change LoRA Rank? — The Two Regimes μA (2026) Splits, and the Limits of That Evidence ♪ Listenable
Two competing beliefs circulate in practice at the same time: that changing the rank in LoRA means you have to re-find the optimal learning rate, and that using 1/r scaling makes the learning rate independent of rank. Th
2026-07-16 · 18 min read #llm#lora#fine-tuning#peft#trainingKafka Diskless Topics (KIP-1150): The Cross-AZ Cost You Trade for Latency — and a Feature Not Yet Shipped
On March 2, 2026, the Apache Kafka community accepted KIP-1150, Diskless Topics — a proposal to make object storage, not broker disks, the source of truth for data, eliminating the cross-AZ replication traffic that is th
2026-07-16 · 16 min read #kafka#streaming#distributed-systems#cloud-nativeKotlin Swift Export Reaches Alpha — It Is Tearing Down the Objective-C Bridge, but Do Not Cross Yet
The biggest tax Kotlin Multiplatform has paid on iOS has been the Objective-C bridge. Kotlin code could not reach Swift without passing through an Objective-C header, and in that crossing, Int? got boxed into KotlinInt,
2026-07-16 · 17 min read #kotlin#kotlin-multiplatform#swift#ios#compilerThe Version After MySQL 9.7 Is 26.7 — How MySQL and MariaDB Tidied Up Their Release Models in Spring 2026 ♪ Listenable
On April 21, 2026, MySQL 9.7.0 shipped as GA, opening a new LTS line for the first time in almost two years since 8.4 — and that same month, Extended Support for MySQL 8.0 ended, dropping it into Sustaining Support. Then
2026-07-16 · 20 min read #mysql#mariadb#database#versioning#devopsThe goroutineleak Profile in Go 1.26 — Catching Goroutine Leaks with GC Marking
Go 1.26 (released February 10, 2026) added a goroutineleak profile to runtime/pprof as an experimental feature. The idea is to recycle the GC's marking phase — mark using only runnable goroutines as roots, then report as
2026-07-16 · 17 min read #go#goroutine#garbage-collection#profiling#debuggingllama.cpp Comes to the Browser — The Ceiling a WebGPU Backend Measured Across 16 Devices
LlamaWeb, the WebGPU backend for llama.cpp released by a UC Santa Cruz team in May 2026, left behind the widest dataset to date measuring browser LLM inference — 16 devices across 8 vendors. The results cut both ways. It
2026-07-16 · 16 min read #webgpu#llm-inference#llama-cpp#on-device-ai#browserMove to Ambient and EnvoyFilter Goes Silently Ignored — The Gap Istio 1.30's TrafficExtension Closes, and What It Doesn't
The real obstacle blocking migration to ambient mode isn't resources — it's extensibility. Istio's official migration doc states outright that EnvoyFilter is not supported on waypoints, that it will be "silently ignored"
2026-07-16 · 15 min read #kubernetes#istio#service-mesh#envoy#ambient-meshThe Workload/PodGroup API in Kubernetes v1.36 — Gang Scheduling Is Moving Into kube-scheduler ♪ Listenable
Gang scheduling for AI training and batch workloads has so far been the job of external schedulers like Volcano or Kueue, but Kubernetes has started pulling this capability into its core. v1.35 shipped the Workload API a
2026-07-16 · 19 min read #kubernetes#scheduling#gang-scheduling#distributed-training#draInfluxDB 3 Core's 72-Hour Limit Is Actually a 432-File Limit — The Bill Left by Rewriting Parquet ♪ Listenable
InfluxDB 3 rewrote its entire engine in Rust and put the storage layer on Apache Arrow and Parquet. It is commonly claimed that Core only queries the last 72 hours, but reading the source shows no such code. What actuall
2026-07-16 · 18 min read #database#influxdb#time-series#parquet#storage-engineWhy the Default Vector Index Moved from HNSW to Disk — the Tradeoffs of Elasticsearch bbq_disk ♪ Listenable
As of Elasticsearch 9.4 (2026-05-05), the default index type for float vectors changed to bbqdisk — a disk-based IVF index. In under a year the default moved from int8hnsw to bbqhnsw and then, a second time, to a disk in
2026-07-16 · 23 min read #vector-database#elasticsearch#ann-index#quantization#hnswFerrocene's Certified core — The Compiler Is ASIL D, So Why Does the Library Stop at ASIL B? ♪ Listenable
The claim that Ferrocene's Rust compiler is 'ASIL D certified' is true, but the conclusion people usually draw from it is not. Ferrocene qualified the compiler as an ISO 26262 ASIL D / TCL 3 tool, which means 'you may de
2026-07-16 · 23 min read #embedded#rust#ferrocene#safety#systemsThe Final Field Mutation Warning in JDK 26 — What JEP 500 Actually Changed, and What to Check Now
JDK 26 (GA on March 17, 2026) started emitting a runtime warning, via JEP 500 "Prepare to Make Final Mean Final," for code that reassigns final fields through deep reflection. Ever since JDK 5, Field.set() after setAcces
2026-07-16 · 17 min read #java#jvm#reflection#serialization#migrationIceberg v3 Row Lineage: Row IDs Aren't Stored in the File ♪ Listenable
Apache Iceberg format v3 writes row lineage — a stable identifier for every row — into the spec, and it isn't optional: any v3 table has it on unconditionally. But that value isn't stored in the data file — it's computed
2026-07-16 · 18 min read #iceberg#lakehouse#table-format#data-engineering#sparkDuckDB Gets a Client-Server Protocol — What Quack Changes and What It Doesn't ♪ Listenable
On May 12, 2026, the DuckDB team announced Quack — a client-server protocol built on top of HTTP that lets two DuckDB instances become client and server to each other, so the same database can be read and written concurr
2026-07-16 · 18 min read #database#duckdb#olap#analytics#data-engineering