@g. coleman@Guiber Acosta 👋 Welcome to the Cerbos community. How did you hear about Cerbos?
g
g. coleman
07/11/2022, 7:21 PM
via HackerNews threads where charithe advertised. Very stoked to find and learn!
g
Guiber Acosta
07/11/2022, 7:27 PM
Via medium , currently I'm working in a gql proxy for ABAC model
e
Emre (Cerbos)
07/11/2022, 7:28 PM
Great to hear! Please let us know if there is any info you cannot find online. We are also happy to schedule a workshop if you need help with anything.
g
g. coleman
07/11/2022, 7:30 PM
i'm interested in understanding perf and scale.. any benchmarking topic links appreciated
g. coleman
07/11/2022, 7:32 PM
also still diving into query planner .. trying to grok how far this can/will go. Of particular interest is an Elasticsearch adapter
a
Alex Olivier (Cerbos)
07/11/2022, 10:53 PM
Query Plan: How are you getting on with this? I spoke about this and gave a live demo with the Prisma team that might help you grok things a bit more
https://youtu.be/lqiGj02WVqo?t=4319▾
We don’t today have an adapter for ElasticSearch (would love contributions if you want to!) but here is a one for Node + Prisma https://github.com/cerbos/query-plan-adapters/blob/main/prisma/src/index.ts
The theory is the same, map the AST into the structure needed for your query filter.
Scale: We have some internal testing with 100s and 1,000s of policies and get the same millisecond response times. There are also some users who have pushed beyond this in testing with 100,000+ policies with still no issues to report.
Perf: Whilst we don’t have formal benchmarks, you can see the results from our internal tests in this PR which is the switch away from using a 3rd party to our own in-house decision engine last year with up to 17x performance improvements in certain areas
👍 1
d
david
07/24/2022, 7:35 AM
Sorry to hijack this thread but I was searching Slack for similar answers 🙂
Regarding scale, do you have information about CPU/memory requirements for loading and evaluating 100s, 1000s, 10000s of policies? Im assuming that since there is no data store all the policies are loaded/compiled into memory?
(The perf improvements from the attached PR look very impressive BTW!)