Skip to content

The Performance Analyzer RCA is a framework that builds on the Performance Analyzer engine to support root cause analysis (RCA) of performance and reliability problems for Elasticsearch instances. This framework would conduct real time root cause analysis of such problems using Performance Analyzer metrics. Root cause analysis can significantly …

License

Notifications You must be signed in to change notification settings

rguo-aws/performance-analyzer-rca

Repository files navigation

Performance Analyzer RCA

The Performance Analyzer RCA is a framework that builds on the Performance Analyzer engine to support Root Cause Analysis (RCA) of performance and reliability problems in Elasticsearch clusters. This framework executes real time root cause analyses using Performance Analyzer metrics. Root cause analysis can significantly improve operations, administration and provisioning of Elasticsearch clusters, and it can enable Elasticsearch client teams to tune their workloads to reduce errors.

RCA Overview

The RCA framework is modelled as a distributed data-flow graph where data flows downstream from the leaf nodes to the root. Leaf nodes of the graph represent Performance Analyzer metrics on which intermediate computations are performed. The intermediate nodes can be RCAs or other derived symptoms which helps in computation of the final RCA. The framework operates on a single analysis graph which is composed of multiple connected-components. Each connected component contains one top-level RCA at its root.

Terminologies

DataStream a.k.a FlowUnit: A flow-unit is a data object that is exchanged between nodes. Each unit contains a timestamp and the data generated at that timestamp. Data flows downstream from the leaf nodes to the root.

RCA: An RCA is a function which operates on multiple datastreams. These datastreams can have zero or more metrics, symptoms or other RCAs.

Symptom: A symptom is an intermediate function operating on metric or symptom datastreams. For example - A high CPU utilization symptom can calculate moving average over k samples and categorize CPU utilization(high/low) based on a threshold. A symptom node typically outputs a boolean - presence(true) or absence(false) of a symptom.

Metrics: Metrics are typically served as continuous datastreams to downstream nodes in the graph from Performance Analyzer (PA). The RCA framework can also be extended to pull custom metrics from other data sources.

Components

Framework: The RCA runtime operates on an AnalysisGraph. You can extend this class and override the construct method to build your own RCAs. You should specify the path to the class in the analysis-graph-implementor section of pa_config/rca.conf. The addLeaf and addAllUpstreams helper methods are useful when you define the dependencies between nodes of the graph.

Scheduler: The scheduler invokes the operate method on each graph node in topological order as defined in the AnalysisGraph. Nodes with no dependencies can be executed in parallel. Use flow-units to share data between RCA nodes instead of shared objects to avoid data races and performance bottlenecks.

Networking: The networking layer handles RPCs for remote RCA execution. If a host depends on a remote datastream for RCA computation, it subscribes to the datastream on startup. Subsequently, the output of every RCA execution on the upstream host is streamed to the downstream subscriber.

WireHopper: Interface between the scheduler and the networking layer to help in remote RCA execution.

Context: The context contains a brief summary of the RCA. For example, a High CPU utilization symptom context will contain the average CPU utilization when the symptom was triggered.

Thresholds: Thresholds are static values that must be exceeded to trigger symptoms and RCAs. Thresholds can be dynamically updated and do not require a process restart. Thresholds are user defined and often depend on hardware configuration and Elasticsearch versions. The threshold store supports tags to help associate metadata with a threshold.

Tags: Tags are key-value pairs that are specified in the configuration file(rca.conf). Tags can be associated with both hosts and RCA nodes.

  • RCA nodes are only executed on hosts with the exact same tags as the RCA node. A common use-case of tags is to restrict certain RCA nodes to only execute on the master node.
  • Tags are also used by hosts to find and subscribe to remote datastreams. For example, a cluster-wide RCA running on the master can subscribe to datastreams from all data hosts in the cluster.

Design RFC

RFC

Building, Deploying, and Running the RCA Framework

Please refer to the Install Guide for detailed information on building, installing and running the RCA framework.

Current Limitations

  • This is alpha code and is in development.
  • We don't have 100% unit test coverage yet and will continue to add new unit tests.
  • We have tested and verified RCA artifacts only for Docker images. Other distributions are yet to be built and tested and will be available as a part of the final release.

Contribute

We invite developers from the larger Open Distro community to contribute and help improve test coverage and give us feedback on where improvements can be made in design, code and documentation. You can look at Contributing.md for more information on how to contribute.

Code of Conduct

This project has adopted an Open Source Code of Conduct.

Security Issue Notifications

If you discover a potential security issue in this project we ask that you notify AWS/Amazon Security via our vulnerability reporting page. Please do not create a public GitHub issue.

License

This library is licensed under the Apache 2.0 License.

About

The Performance Analyzer RCA is a framework that builds on the Performance Analyzer engine to support root cause analysis (RCA) of performance and reliability problems for Elasticsearch instances. This framework would conduct real time root cause analysis of such problems using Performance Analyzer metrics. Root cause analysis can significantly …

Resources

License

Code of conduct

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages