bellwood

This release is relatively minor. It introduces self-telemetry into the wavefront sink which, it is hoped, will help us debug an alerting issue seem on some of Postmate's internal time-series.

This resolves #117

bellwood

This release introduces severl long-awaited changes. We now expose the internal routing of cernan to the user via configuration and introduce a prototype filter sub-system which allows on-the-fly reprogramming of cernan by users.

• Resolves #115
• Resolves #45

We're very, very close to the 0.5 milestone.

bellwood

This release contains two key features. The first, and most immediately important, are adjustments to the file watching subsystem. We no longer use filesystem events and prefer to do periodic polling. This resolves overload of the OS notify mechanism on systems with very fast logs. As well, this release includes the start of the scripting filter mechanism though it is disabled.

bellwood

This release includes a DNS modification to the wavefront sink. Previously we would cache the DNS -> IP lookup and this didn't always work out for us in practice.

bellwood

This release contains a bug-fix to the cernan executable. In particular, we were passing the flush-interval config value down through to the wavefront sink rather than the sink's own bin_width value.

bellwood

0.4.3 had a bug in which the flush timer thread was not 'run'. This
represented a change to the API for all sources and this was missed
for the flush timer and file server.

bellwood

This release addresses operational and correctness issues. In
particular:

• we've added significantly more logging to cernan (#136)
• we now fast-forward mpsc receiver (#137)
• metric bins are now scalable, qos is removed (#134)
• file source can watch directories of files (#132)
• federation receiver does not panic on hangup (#133)
• metrics now have an associated algebra (#131)
• metrics carry their own metadata (#129)

Signed-off-by: Brian L. Troutwine blt@postmates.com

bellwood

This release incorporates several changes in many areas. In particular:

• [EXPERIMENTAL] cernan is now able to forward on to other cernan instances
• [STABLE] all metrics carry their timestamps, including histograms / timers
• [STABLE] QOS sampling is now done by random choice in the flush interval
• [EXPERIMENTAL] more file source changes have landed, improving use

The main show here is the sampling. In particular, the QOS sampling now
is done such that, if QOS is Q and the flush interval is F then there will
be no more than ceil(F/Q) points in the interval F. If there are less than
ceil(F/Q) points in F then all points will be shipped. If there are more,
the sampling will be done on a random basis to avoid introducing systematic
artifacts into the time series.

bellwood

This release puts timers and histograms onto the same time-accurate
footing as other metric kinds, very useful when the underlying data-
source is sparse, as has been observed. Some configuration changes
have also been made--in a fully backward compatible fashion--but
these are secondary to the time-accuracy modifications.

Signed-off-by: Brian L. Troutwine blt@postmates.com

bellwood

This release incorporates the final work to move away from std's in-memory channel, freeing up cernan to emit toward slow sinks--firehose--and ingest more data for analysis that will fit into memory. Performance remains on par with the previous cernan series and has been improved in regards to statsd/graphite parse speed. A variety of filesystems have been tested and it is believed that cernan will work correctly no matter what goofy filesystem you have it queue toward.

cernan is now well arranged to allow in-flight analysis and manipulation of data, as well as forward off-system to itself for aggregation should the present system be particularly small but heavily telemetered.