Ecommerce ops · Production

Space-efficient ML feature stores using bloom filters: a Zalando engineering benchmark

The problem

Conventional key-value-store-based ML feature stores become very large and impose significant challenges: network calls add 2–10ms of latency per lookup, distributed databases with strict performance requirements are expensive to host, backfill operations become very costly, and multiple feature lookups per request become prohibitively expensive under strict latency budgets.

Workflow diagram · grounded in source
1
User request triggers lookup
trigger
“When a request is made to the recommender system, a query is made to this key-value store using the user ID”
2
In-process feature store query
integration
“read a much smaller dataset into memory, and query that as a feature store from within the process”
3
Click prediction
ai_action
“a logistic regression classifier was used to predict clicks”
Reported outcome

A bloom-filter-backed compressed feature store achieves the same click-prediction classification performance (AUC~=0.7997) as a conventional key-value store while using only 3% of the memory, with no detectable throughput overhead.

Reported metrics
Compressed vs uncompressed feature store memory footprint3%
click-prediction AUC at 3% memoryAUC~=0.7997
uplift retained at 40MB state size90%
State size at 90% uplift as fraction of uncompressed0.3%
Show all 6 reported metrics
compressed vs uncompressed feature store memory footprint3%
click-prediction AUC at 3% memoryAUC~=0.7997
uplift retained at 40MB state size90%
state size at 90% uplift as fraction of uncompressed0.3%
prediction throughput (all variants)20-22k predictions per second per core
latency added by external feature store network call2-10ms
Reported stack
RedisBloom-Filter
Source
https://engineering.zalando.com/posts/2021/10/space-efficient-machine-learning-feature-stores-using-probabilistic-data-structures.html
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

A bloom-filter-backed compressed feature store achieves the same click-prediction classification performance (AUC~=0.7997) as a conventional key-value store while using only 3% of the memory, with no detectable throug…

What tools did this team use?

Redis, Bloom-Filter.

What results were reported?

Compressed vs uncompressed feature store memory footprint: 3%; click-prediction AUC at 3% memory: AUC~=0.7997; uplift retained at 40MB state size: 90%; State size at 90% uplift as fraction of uncompressed: 0.3% (source-reported, not independently verified).

How is this ecommerce ops AI workflow structured?

User request triggers lookup → In-process feature store query → Click prediction.