Evolving Pinterest's Embedding Retrieval Platform (8 minute read)
Pinterest's Manas retrieval platform serves billions of embeddings across Home Feed, Search, Related Pins, Ads, and Notifications. Quantization is cutting serving cost by 20–30%, SSD-backed ANN experiments reduce memory and CPU needs, and multi-embedding retrieval moves beyond single-vector two-tower matching toward richer candidate scoring.
|
Diskless Kafka: What Happens When Brokers Stop Owning the Data? (16 minute read)
Diskless Kafka keeps the Kafka API while moving durable payload storage from broker-owned logs to object storage and strongly consistent metadata. The tradeoff is not just cheaper retention: batching, offset assignment, recovery, compaction, latency, and feature support all change, making diskless topics a per-workload storage choice rather than a universal Kafka replacement.
|
|
The unbearable lightness of one more index (11 minute read)
Coding agents now produce competent PostgreSQL schemas, but they tend to stack indexes onto hot operational tables. In a measured support-ticket workload, extra indexes nearly doubled update time and WAL while increasing VACUUM work. Index review needs write-path evidence, HOT-update awareness, and usage counters, not just fast read plans.
|
|
118 million queries per second on Neki (9 minute read)
PlanetScale set out to run 1 million queries per second against Neki, its new sharded Postgres, hit that on 5 shards almost immediately, then kept going until 512 shards were serving 118 million queries per second across 1.22 PiB of data. The workload is deliberately narrow: a single-shard point select fetching one row by primary key with no writes, joins, or cross-shard queries, which is what makes the scaling curve clean.
|
Now everyone can put data to work (3 minute read)
ChatGPT Work's Data agent connects to approved company data and business definitions, letting employees investigate questions and build shareable dashboards using plain language. It respects existing permissions, works with major data and BI platforms, and can recommend or carry out approved follow-up actions.
|
Why DuckDB 2.0 is faster (15 minute read)
DuckDB 2.0 speeds up S3 queries by 2-3x through asynchronous I/O, dramatically improves deep recursive queries, and makes semi-structured VARIANT data smaller and faster than JSON. It also adds practical SQL features including triggers, nested schemas, and data-modifying CTEs.
|
|
Your AI Adoption Lift Is a Selection Effect (17 minute read)
Comparing AI adopters with non-adopters usually exaggerates impact because more engaged customers are also more likely to adopt. Instead, estimate causality using variation that customers did not choose, such as an eligibility threshold, while clearly limiting conclusions to the population and decision it measures.
|
|
|
|
|
