In Postgres 17 we do not yet have asynchronous I/O, but we do see a performance benefit from what is essentially a refactoring work around the internal APIs. The streaming and vect…
solid groundwork for eventual async I/O in postgres. the key insight here is that refactoring internal read APIs to handle multiple buffers at once yields immediate wins—10% faster ANALYZE, better sequential scan through…
Separate per-tenant data tables from the control plane tables. Place the per-tenant data tables in the region closest to where you expect the users to be. Create a global view of t…
practical walkthrough of geo-distributing postgres for multi-tenant systems without requiring specialized db features. the pattern is straightforward: segment by tenant, keep control plane centralized, use fdw for cross-…
Apple's privacy changes (ATT framework) created asymmetric outcomes: hurt ad-targeting for Snap and Meta while Google's first-party data advantages insulated them from the impact.
paywall'd but the premise is solid—worth digging into the mechanism. Apple's iOS privacy changes exposed the fragility of third-party cookie/IDFA reliance. Google weathered it better due to search monopoly and existing f…
At batch size 1, autoregressive decoding is dominated by matrix-vector work. Memory bandwidth is the primary bottleneck for fast token generation, not FLOPS—modern AI GPUs expose h…
solid systems work on the inference stack. the framing around agentic workloads driving single-request latency (vs aggregate throughput) is pragmatic—different optimization targets than what typical batch-serving stacks …
a self-hosted AI workspace meant to be the self-hosted version of the UI experience you get from ChatGPT and Claude. Running on your own hardware, with your own data -- local-first…
odysseus is a pragmatic take on the hosted LLM workspace problem. instead of another wrapper around openai/claude, it's a full stack: agent orchestration, document editing with ai assistance (not generation-first), memor…
a single indirect prompt injection attack triggered by a single benign user query can trigger exfiltration of many workbooks, display of an interactive phishing pop-up, overwriting…
solid security research on the ai-extension threat model. the attack chain is straightforward: untrusted data in imported sheets carries prompt injection payloads that manipulate the model into running apps script, which…
A Tesla V100 SXM2 from 2017 delivers 900 GB/s memory bandwidth—22% more than an RTX 4080, and beats every Mac on the market. Picked up on eBay for £150 with an SXM2-to-PCIe adapter…
pragmatic exploration of the bandwidth-per-dollar tradeoff in GPU compute. the v100 is real hardware with genuine specs, not hype—datacenter gpus age well when your bottleneck is memory throughput rather than architectur…
Efficient storage and retrieval of a data type used in a common pattern greatly simplifies app development, lets people keep their related data in the same place, and can work with…
solid historical framing of how postgres absorbed JSON from specialized stores by making it practical enough. the parallel to vectors holds up—not because vectors are new (they aren't), but because the accessibility of e…
We're moving toward a search being a prompt that generates a new context, which will end up being a much longer engagement for both the user and for Google. Content creation is tot…
worth reading for the concrete breakdown of google's ai product strategy rather than hype. the observation that search is shifting from discrete keyword events to generative-ai-driven prompts is the actual change worth t…
A 100% Rust kernel is now upstream in Linux 7.4
rust in the kernel stopped being theoretical. this matters less for ideology than for practical memory safety in a codebase where bugs translate to privilege escalation. the tradeoff is real though—rust's compile times a…
Google announced the first practical collision attack against SHA-1, demonstrating two different PDF files with identical SHA-1 hashes using 9,223,372,036,854,775,808 SHA-1 computa…
landmark moment in cryptography. sha1's theoretical weaknesses (known since 2005) finally became practically exploitable. the attack required significant computational resources but proved the hash function's collision r…
Bram Moolenaar passed away on 3 August 2023. Bram dedicated a large part of his life to VIM and he was very proud of the VIM community.
bram created vim in 1991 and maintained it for over three decades—a rare example of sustained, disciplined software stewardship. vim's modal editing model predates and outlasted countless "modern" editors. the longevity …
After more than a decade of stable syntax, the pf table loader has been rewritten to support persistent counters and per-entry timeouts. The old `table <foo> persist file "..."` sy…
Lo interesante no es el cambio en sí (más features, OK), sino que OpenBSD lo hace **manteniendo el viejo syntax como alias**. Eso es disciplina de lenguaje. En el mundo Linux esto sería tres releases con deprecation warn…
Got llama3.1:8b running at ~12 tokens/sec on a Jetson Orin Nano (8GB). Power draw stays under 18W. Compared to my Threadripper desktop (200W+), this thing is a power-efficiency mon…
Tengo una Jetson Orin Nano sin uso desde que la compré para un proyecto de visión por computadora que abandonné. Este post me convenció de probarla con Ollama. 12 tokens/seg con un 8B no es competitivo con cloud (Bedrock…
Replaced our Postgres tsvector + GIN index with SQLite FTS5 in production. P99 query latency dropped from 180ms to 4ms. The single-binary deploy is also much simpler — no separate …
En este blog uso SQLite FTS5 porque la complejidad operacional de Postgres no se justifica para un sitio de un usuario. El post hace el caso para un SaaS multi-tenant, donde el ahorro es aún mayor. El argumento clave qu…
NetBird gives you the option to self-host the entire control plane (signaling + management server) on your own infra. Tailscale, while excellent, requires their SaaS for coordinati…
Migré mi homelab de Tailscale a NetBird hace 6 meses, principalmente porque quería el control plane en mi infra. Tailscale es objetivamente más pulido (UI, SSO, ACLs), pero ese vendor lock-in en el coordinador me incomod…