This message was deleted.
# general
s
This message was deleted.
f
We have a few JMH benchmarks checked into the repo. Besides that the rest of our performance-related activities aren't in open source. We have some environment where we can generate synthetic traffic (been trying to convince @Min Huang to open source the synthetic traffic generation app). But the ultimate test is the production workloads... we take flamegraphs from clients or clusters that are struggling and then optimize the hot paths.
e
Makes sense
do you have some client to run benchmarks ? especially for the read workflow
d
Storing bookmark to https://github.com/openjdk/jmh
m
@Felix GV I shared the same vision; just need to find the time to move them, and test in non-LI environment.
🤝 1
Read traffic might be easy; as for write traffic, they currently depends on Hadoop and Samza; I am not sure how to generate those traffic without spinning up Hadoop and Samza first.
f
Yep, let's start with reads then.
Then maybe we can throw in @Nisarg Thakkar's online producer in there too, to have synthetic R/W traffic from a single app, no Samza or Hadoop needed.
I assume he'll probably want to add it there anyway.
e
Now we also have Pulsar to generate load 🙂 At DataStax we have plenty of tools to generate Traffic on Pulsar
we will probably start filling in a Venice cluster from Pulsar and then perform random reads to Venice and measure latency
f
Looking forward to hearing your feedback on that! There are some important RocksDB knobs that can be tuned for cost vs performance. Happy to dive into it more when you get there.