Cache Adapter Questions about ruby-client

[patricioe]

So given we want to test all public traffic... should the number here be Float::INFINITY or something?

Feel free to reach out to [email protected] so we can privately ask some numbers around how many impressions/feature flags evaluations you may producer and we can figure out the right number.
If you produce more than the queue size, usually impressions are dropped and we show a warning.

[patricioe]

@metaskills Ken, quick follow up here.

When ruby operates in consumer mode using redis, impressions_queue_size is not used at all.

[metaskills]

FYI, I was reading the detailed doc here (https://github.com/splitio/ruby-client/blob/master/Detailed-README.md) when I thought of this questions. I think maybe most people use a Redis backend or maybe a lot use memory. I just wanted to be sure I understood the differences when testing the product at scale.

[maksimf]

What is stored in memory? Splits & Envs only? Per user treatments?

When using MemoryAdapters we store splits, segments, impressions and metrics data in memory, here you can see Repository objects. When we need to store something we use Repository for that, whereas Repository uses one of available adapters: :memory or :redis. Memory adapters diverge further, currently there are 2: MapAdapter and QueueAdapter.

Will the memory grow unbound or does it get collected?

Memory do not grow unbound, though it's not recommended to use memory adapters for a larger number of splits/segments/impressions/metrics (i.e. "descent" amount of web traffic) and this is why:

• for splits and segments we don't remove them from memory at all, only add new and update old ones
• for impressions and metrics we do remove them from memory after we bulk sent them to the Split API, but if you have a lot of impressions it can become a bottleneck: much more impressions coming in, then we remove. You can tweak size of the impressions queue and "refresh_rate" (how often impression queue get cleared), but it might not always help.

All things considered I'd recommend using Redis as adapter for scalability and memory for simplicity (faster and easier to set up).

Did I answer these questions? If not, please feel free to ask more.

[metaskills]

You did answer that really well, thanks! Going to move forward with redis as the store. A few more questions then.

• Can we assume all local data in redis if lost would rebuild from Split.IO? Not likely, just curious in case the Redis cluster was setup to act like an LRU.
• Reading the docs/source for impressions_queue_size. I feel like the best setting would be -1?
• When using redis, what mode would all app servers use? I'm thinking standalone the default is best because it would appear to mix both producer and consumer, is that true? If so, would it be recommended to spin up some other additional process on one of n servers to be a producer? Just looking for an ideal canned setup for a single & suite of Rails apps using redis.

Thanks again!

[maksimf]

Can we assume all local data in redis if lost would rebuild from Split.IO? Not likely, just curious in case the Redis cluster was setup to act like an LRU.

Yes and no. Data which comes to you from Split API (i.e. splits and segments) will be restored, but impressions and metrics would be completely lost (since they are generated on your server(s) and then sent to Split API).

Reading the docs/source for impressions_queue_size. I feel like the best setting would be -1?

Not really, maybe it's not spelled right in the docs, but if you set impressions_queue_size to -1 it would disable impressions at all

When using redis, what mode would all app servers use? I'm thinking standalone the default is best because it would appear to mix both producer and consumer, is that true?

Yep, it's true

If so, would it be recommended to spin up some other additional process on one of n servers to be a producer?

Tough one, it completely depends on your infrastructure design, it's hard for me to give some abstract advice here.

[metaskills]

Not really, maybe it's not spelled right in the docs, but if you set impressions_queue_size to -1 it would disable impressions at all

So given we want to test all public traffic... should the number here be Float::INFINITY or something?

[patricioe]

When using redis, what mode would all app servers use? I'm thinking standalone the default is best because it would appear to mix both producer and consumer, is that true?

You will want to have on producer (we call it Split-sync and it's written in Go. https://docs.split.io/docs/split-synchronizer) and many Ruby consumers.
If you use many standalone ruby sdks, they will all write to redis.

[maksimf]

@metaskills One more thing, I have to correct myself: we do delete splits from the cache (memory or redis) when they become archived (you can mark split archived in the UI)

[maksimf]

@metaskills Should we close this issue? Or can I answer anything else?

[metaskills]

Closed. Thanks again!