The general premise of this release is to make metrics more flexible within Hystrix. See https://github.com/Netflix/Hystrix/wiki/Metrics-and-Monitoring for a deep dive on the new metrics architecture. The high-level approach is to model metrics directly as a stream, so that Hystrix metrics consumers may aggregate the metrics as they wish. In 1.4.x and prior releases, HystrixRollingNumber and HystrixRollingPercentile were used to store aggregate command counters and command latencies, respectively. These are no longer used.
Instead, new concepts like HystrixCommandCompletionStream are present. These may be consumed by a rolling, summarizing data structure (like HystrixRollingNumber), or they may be consumed without any aggregation at all. This should allow for all metrics processing to move off-box, if you desire to add that piece to your infrastructure.
This version should be backwards-compatible with v1.4.x. If you find otherwise, please submit a Hystrix issue as it was unintentional.
This version also introduces new metric streams: (configuration and Utilization) have been added, along with a request-scoped stream.
Archaius is now a soft-dependency of Hystrix, so you can supply your own configuration mechanism.
Some known semantic changes:
- Latencies for timeouts and bad-requests are now included in command latency
- Latency distribution percentiles are now calculated with HdrHistogram library and don't have a max number of elements in the distribution
- Previously, HealthCounts data allowed reads to see the value in the "hot" bucket. (the one currently being written to). That does not happen anymore - only full read-only buckets are available for reads.
- Bucket rolling now happens via Rx background threads instead of unlucky Hystrix command threads. This makes command performance more predictable. User-thread latency is now practically indistinguishable from command latency.
Artifacts: Maven Central, Bintray