My understanding of the produce message conversions metric is that the metric is specifically for message format up-conversion which is independent of TLS being enabled.
It’s also worth noting that the zero-copy sendfile is not used on the produce path, only the fetch path. On Thu, Feb 23, 2023 at 4:56 AM Sergio Daniel Troiano <sergio.troi...@adevinta.com.invalid> wrote: > Hey lads, > > I was thinking about message conversions and zero copy, as we know TLS will > break the zero copy as the encrypt/decrypt happens in userland, correct? > > For example ProduceMessageConversionsPerSec will be triggered when a > message must be converted (for example compressing at topic level). > > Should we consider the TLS layer a conversion as well? My concern is we are > "hiding" this metric, so let's suppose I have my cluster using TLS (it > means zero-copy is already "broken") then I activate the compression at > topic level. > Now I will see the values in ProduceMessageConversionsPerSec as the > conversion is happening (tracking the compression at topic level) but the > reality is it was already happening before because of the TLS, so in terms > of performance it's a kind of illusion as it seems the conversion is not > costing any extra CPU cycles. > > If you think it is a good idea maybe we could include the count on the > ProduceMessageConversionsPerSec when the TLS is enabled, what do you think? > > Thanks in advance. > Sergio Troiano >
