I am a Principal Software Engineer (a role I stepped into at the beginning of 2026) based in Vietnam ๐ป๐ณ. I have a deep passion for technology, especially AI.
Principal Software Engineer in Vietnam, building applied AI and the systems underneath it.
Open source I maintain:
- apex-skills: Agent Skills + a deterministic context engine for AI coding agents
- claude-code-otel: local OpenTelemetry stack for Claude Code
- spring-batch-cronjob: distributed batch on Java 25 / Spring Boot 4
https://t.co/e9LNpBqBo2
Open to Principal AI Engineer and Technical Lead roles.
Source: "15 million web servers on one VM - smoltcp, DPDK, AF_XDP" by Artem (kapteko), Nov 18, 2026. Figures are from the author's proof-of-concept benchmark.
https://t.co/klyviL3l0m
HN discussion: https://t.co/udjUar4Zni
Someone fit 15M HTTP listeners on one VM with 10GB RAM and 1 core. Fun stunt, real platform lesson.
The wall was never throughput. It was idle cost per tenant. A process: ~1.5MB. A kernel TCP socket: 3,384 bytes before any traffic. Kernel stack OOMs near 2.5M servers. Userland smoltcp: ~620 bytes idle, past 16M.
The bill moves, not vanishes: you now own TCP correctness, TLS and the tools ss/tcpdump gave you.
Density platforms will compete on idle bytes per tenant.
Source: Zerodha Tech Blog, "Distributed, high-performance blob storage on SQLite" (benchmark numbers are their own, run on tmpfs): https://t.co/wUFtMYjah2
HN discussion: https://t.co/Z370apQNCo
Zerodha's SQLite blob store shows where the bottleneck really was. 1 to 16 shards, same total resources (their tmpfs benchmark): writes 7,740 to 25,324 ops/s, reads 67,128 to 70,884. WAL readers were never the problem. The single write lock was.
Prod said it too: p99 writes under 5 ms spiked to 30-50 ms when TTL cleanup grabbed that lock with one big DELETE.
What I'd copy: one writer queue per shard, and every write goes through it, janitors included.
Source: NVIDIA SkillEvaluator, an open-source tool that checks whether agent harnesses load a skill, runs quality and security scans, and compares agent runs with and without the skill. https://t.co/oJHZSpvlDk
Agent skills are quietly becoming a supply chain, and most teams test them the way we tested shell scripts in 2010: not at all.
NVIDIA's SkillEvaluator gets the order right. First: will the harness even load it? A description a few characters over a 1,024 limit and the skill silently never exists. Then static and security checks. Then the only number that matters: task results with the skill vs without.
If you can't show lift, it's just more context you pay for.
Source: Tech Stackups, "Quad9 Is Leaving France: Why DNS Resolvers Are the Wrong Place to Fight Piracy" https://t.co/sk5OmhKGTO
HN discussion: https://t.co/Sv1VBE8X6L
France is about to run a live experiment on what happens when you enforce policy at the wrong layer of the stack.
On November 24, Quad9, the Swiss non-profit behind 9.9.9.9, stops answering DNS queries from France. Not because of cost or traffic. A Paris court put a 1,000 EUR/day penalty on it for refusing to block 58 pirate tennis-streaming domains, and roughly 20 similar cases are still pending. Quad9 says it cannot survive the accumulation.
The engineering problem is more interesting than the legal one.
1. A resolver has no concept of "a French user."
Quad9 keeps no user records by design. The order effectively demands per-jurisdiction answers. The only tool for that is GeoIP, which is a probabilistic guess about an address, not a fact about a person. Quad9 already tried geo-blocking in Germany during the Sony Music case. Sony's lawyers used a VPN to reach the content "from Germany" and the court called it non-compliance. So the real options are: lie to the whole planet, build a location system that gets judged on its worst case, or exit.
2. The safety feature became the evidence.
The court pointed out that Quad9 already filters malware domains globally, so it can obviously drop names. That is a design lesson worth saying out loud: any filtering primitive you ship for safety will eventually be read as a general-purpose control plane by someone with a court order. If you build the hook, plan for who else will want to call it.
3. The block never touches the data plane.
DNS holds pointers, not content. The stream stays up for anyone on another resolver, DoH, or a VPN. Arcom's own numbers: 20% of French users already use an alternative resolver and 29% use a VPN. Arcom reports 12,606 sports domains blocked since 2022, 6,496 in 2025 alone (up 71%), while piracy fell 4% that year. Block counts grow far faster than the problem shrinks, because every mirror is just a new name to chase.
4. Compliance cost is a moat.
Cloudflare satisfied the court by acting at its CDN and kept 1.1.1.1 out of it. OpenDNS left France in 2024. Quad9 only has a resolver, so it leaves. Any rule with a fixed legal and engineering cost per jurisdiction filters out small operators first. And it is spreading: Italy fined Cloudflare 14.2M EUR over 1.1.1.1, and an EU Parliament study proposes 30-minute blocking across resolvers, VPNs and CDNs.
5. The failure is silent.
When a resolver disappears, nobody sees an error. Queries time out and "the internet is down." RFC 8914 already defines Extended DNS Error 16, "Censored," so a client can tell a court-ordered block from a broken network. Nothing in the French orders requires it.
The takeaway: enforcement belongs at the layer that owns the thing being enforced. Content lives at the host. Pushing the obligation down to the last stateless hop that cannot fight back gets you global overblocking, fewer operators, and the same amount of piracy.
The practical part: grep your configs. Routers, DHCP, VPN profiles, Pi-hole upstreams, base images and embedded devices with 9.9.9.9 or 149.112.112.112 hard-coded. If you have users in France, a dependency you never think about has a hard deadline next Tuesday.
Source: Wiz Research write-up on CVE-2026-123633 and CVE-2026-123007. Fixed in NGINX 1.30.7 stable and 1.31.9 mainline; if you cannot upgrade yet, disabling HTTP/2 is the stated workaround.
https://t.co/vL6c0zebrY
HN discussion: https://t.co/SImCUSYMiu
The new Nginx HTTP/2 RCEs are not parser bugs. They are bookkeeping bugs.
One: a connection stays on the "safe to evict" list after a request starts. Under pressure, Nginx drains it mid-request and keeps using freed memory.
Two: a cancelled stream refunds flow-control credit with no cap. The counter passes the max and an unsigned bounds check wraps.
Derived state drifts. Recheck invariants where state changes, not just at the door. Eviction and cancel paths run least and fail worst.
Source: ExitPath, an open-source guide that grades how hard it is to move off Slack, Notion, Google Workspace and Microsoft 365: https://t.co/RLjUZGAy5u
HN discussion: https://t.co/NQ1lvcNVtu
Contrarian take on SaaS lock-in: the data is the easy part to move.
A new open-source guide grades exits from Slack, Notion, Google Workspace and M365. Teams scores as the hardest. Docs and files export fine. What does not move is behavior: free/busy lookups for rooms, SSO wiring, bots, webhooks, the permission model people learned by accident.
Want a real exit plan? Inventory integrations and workflows, not tables. An export button is not an exit.
Most "add vector search" designs are really sync pipelines: copy the lake into a vector DB, then reconcile two copies forever.
Lakegres puts a clustered, quantized index in Hudi/Iceberg table metadata, updated by the same commits that write rows. Their own numbers on 247M vectors: 63.5 ms warm p50, recall@10 0.95.
The lesson: an index is derived state. Commit it with the data, not in a second system you keep in sync.
Source: https://t.co/GeWAuNFxEF's Java4Shell disclosure, covering the affected CPUs, the JDK versions, and both mitigation flags: https://t.co/D57zsQ784V
HN discussion: https://t.co/hboHP9pNDC
Java4Shell is the kind of bug a fleet audit misses.
Same JDK, same container image, two outcomes: exploitable on hosts with AVX-512 VBMI (Ice Lake, Zen 4+), not on ARM. Memory-safe language, hand-written assembly underneath.
What I'd change:
- record CPU flags next to the image digest
- treat JVM intrinsics as attack surface
- don't blanket-roll -XX:UseAVX=2 as the mitigation. ARM JVMs won't start with it. Mixed fleets need per-arch flags, or just patch.
Source: Cerebras, "Cerebras launches disaggregated inference" (Nov 17, 2026). The 5x throughput-per-megawatt figure is their own production measurement on a GPT-OSS-120B endpoint, compared with a CS-3-only setup. The post also describes their KV Portal, which moves the KV cache between prefill and decode over RDMA. https://t.co/REfKhaM7qn
HN discussion: https://t.co/1Tk2C21eWl
Cerebras now runs prefill on AMD GPUs and decode on its wafer chips in production. Their number: 5x throughput per megawatt vs its chips doing both, same generation speed.
The lesson isn't the chip. Prefill is compute-bound, decode is memory-bound. One pool doing both wastes whatever the other phase doesn't need.
Once you split them, the KV cache handoff becomes your hottest internal API. Budget, monitor and version it like one.
Source: Alibaba's Open Code Review repo. The README explains the hybrid design (deterministic file selection, bundling, rule matching and comment positioning, plus an LLM agent) and has their own benchmark against Claude Code on the same model (AACR-Bench, 200 real PRs from 50 repos). The numbers are self-reported. https://t.co/VeHF6kpyC3
Alibaba open-sourced the AI code reviewer it ran internally for two years, and the design matters more than the model.
File selection, bundling, rule matching and comment positioning are plain code. The LLM only does the judgment.
Their claim: same model as a general agent, higher precision, ~1/9 the tokens, lower recall on purpose.
Right trade. A reviewer that cries wolf gets muted in a week. Put coverage in code, not in the prompt.