FAQ

Why are util-stats in the com.twitter.finagle.stats package?

This is a historical artifact that reflects how the code was originally written as part of Finagle but turned out to not actually depend on anything Finagle-specific. The code was moved to util in order to allow usage without depending on finagle-core. To ease the transition, the package namespace was left as is.

Why were there two(!) different implementations?

This is somewhat historical, as early on in Twitter, Ostrich was created for Scala developers. finagle-stats was written quite a bit later in order to improve upon its implementation in terms of allocation efficiency and histogram precision.

Why is NullStatsReceiver being used?

NullStatsReceiver is used when no metrics implementation is found on the classpath. The section on choosing the stats implementation covers how to pick an implementation for your application. Twitter runs its production services with finagle-stats and it is available on Github and Maven Central.

Why migrate to finagle-stats?

The finagle-stats library is designed to replace Ostrich which was formerly supported by Twitter. The comparison table below illustrates the advantages of finagle-stats over Ostrich.

Ostrich

finagle-stats

Allocations per Stat.add

0 bytes

0 bytes

Highest histogram resolution

p9999

p9999

Sampling error

±5%

±0.5%

finagle-stats allows you to visualize and download the full details of Stat histograms. This is available on the /admin/histogram admin page.

Are metric values reset or latched after an observability collection?

Stats are latched by all two implementations. This means that for each observability collection window, you will see that window’s distribution of values.

There is no notion of latching/resetting a Gauge, as it is always a discrete point in time value. This is the case across all three implementations, except for CounterishGauges, which creates a gauge that models an unlatched counter.

Counters have different behavior depending on which implementation is being used. This in turn has an impact in how you write your queries for your dashboards and alerts. Both Commons Metrics [1] and Commons Stats are not “latched”, which means that their values are monotonically increasing (assuming you never call incr() with a negative value). This implies that you use CQLs rate function if you want to see per minute counts. Ostrich’s Counters are latched when they are collected (think of this as being a “reset”) which means that you should not need to use rate in order to see per minute values.

[1] Note that by using finagle-stats and setting the flag com.twitter.finagle.stats.useCounterDeltas=true you can get per-minute deltas in Commons Metrics.