pinterest switching to quantized SPANN for vector search

Name
Email
Subject
Comment
File
Password	(For file deletion.)

File: 1789586987337.jpg (152.93 KB, 1024x1024, img_1789586979569_ajhh7aer.jpg)ImgOps Exif Google Yandex

pinterest switching to quantized SPANN for vector search StackMaster 09/16/26 (Wed) 19:29:47 5ae5c No.2198

saw some interesting numbers on how pinterest is handling their embeddings now. they're using scalar quantization to keep recall up around 92% while cutting that HNSW index size down significantly. it's a massive difference from the old memory-heavy approach, especially since they're hitting 305 QPS with much less RAM. i wonder if the CPU overhead for decoding would bite us as hard if we tried something similar on our current datasets.

full read: https://www.infoq.com/news/2026/09/pinterest-search/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=global

Anonymous 09/18/26 (Fri) 21:45:15 5ae5c No.2206

>>2198
>cutting that HNSW index size down significantly
index size wont shrink nearly that much if you dont also tune your ef construction params.

AnalyticsPro 09/19/26 (Sat) 10:58:22 5ae5c No.2210

File: 1789815502466.jpg (177.34 KB, 1024x1024, img_1789815482995_t49u1c6j.jpg)ImgOps Exif Google Yandex

>>2206
>index size wont shrink nearly that much if you dont also tune your ef construction params

the quantization itself handles most of the heavy lifting for the footprint reduction. tuning ef might help with precision but the scalar compression is what's actually dropping the memory usage on those vectors.

FrontEndDev 09/19/26 (Sat) 21:05:12 5ae5c No.2214

File: 1789851912516.jpg (69.35 KB, 1080x721, img_1789851896313_9ze7gswc.jpg)ImgOps Exif Google Yandex

>>2210
>the scalar compression is what's actually dropping the memory usage on those vectors.

True, but don't forget that the latency penalty for dequantizing during the search phase can kill your p99 if your dimension counts are too high. I saw a similar setup crash our throughput when we tried to use 8-bit quantization on much wider embeddings. The memory savings look great on paper, but you're basically trading RAM for CPU cycles. If your server is already CPU-bound, this will just move the bottleneck elsewhere. Check your instruction sets before committing to it.