Hmm, we do use the realm tag in our metric publishing. I understand the
concern re: cardinality. Maybe we can support filtering metrics that have
realm and support another metric without realm?

On Mon, May 19, 2025 at 12:24 PM Dmitri Bourlatchkov <di...@apache.org>
wrote:

> Removing realm_id from metrics tags makes sense to me (to avoid high
> cardinality).
>
> If we need to have insight into load differences from realm to realm, it
> might be preferable to introduce metrics dedicated to that rather than
> increasing the cardinality of every endpoint metric.
>
> Cheers,
> Dmitri.
>
> On Thu, May 15, 2025 at 3:30 PM Alex Dutra <alex.du...@dremio.com.invalid>
> wrote:
>
> > Hi all,
> >
> > I would like to suggest removing the "realm_id" metric tag entirely.
> >
> > My concern is that this tag has the potential for high cardinality, which
> > is generally considered a bad practice when dealing with metrics. High
> > cardinality can lead to performance issues and increased memory usage.
> >
> > Granted, the default realm resolver in Polaris is tailored for just a
> > handful of realms, but nothing prevents users from declaring hundreds of
> > realms.
> >
> > I believe we can still effectively monitor Polaris servers without this
> > specific tag, since the realm ID is also propagated in traces emitted by
> > Polaris. Tracing is a much better fit for high-cardinality domains.
> >
> > I'm open to discussing this further; a potential alternative would be to
> > introduce a flag to disable this specific metric tag, but I feel like
> > removing it would be a much cleaner approach.
> >
> > Let me know your thoughts.
> >
> > Thanks,
> >
> > Alex
> >
>

Reply via email to