Skip to content

Latest commit

 

History

History
40 lines (26 loc) · 2.41 KB

PREDICTION.md

File metadata and controls

40 lines (26 loc) · 2.41 KB
Raw

Online Prediction

MeshInsight relies on critical path, sidecar proxy configurations, and workload characteristics to generate overhead prediction. We provide two options for you to provide these information

  1. We have integrated MeshInsight with CRISP, a critical path extractor developed by Uber. MeshInsight also tries to extract the proxy configuration from Kubernetes deployment files if possible.
  2. You can also provide the critical path, sidecar proxy configurations, and workload characteristics in JSON format (see below)

Option 1

To use the CRISP option , you need to provide a few parameters (in yaml format) listed below.

CRISP:
  # path of the trace directory (traces must be in JSON format)
  TRACE_DIR: meshinsight/predictor/CRISP/example1
  # service name and operation name are in Jaeger terminology
  SERVICE_NAME: service-a
  OPERATION_NAME: ping-receive
  # Should the service and operation be the root span of the trace
  ROOT_TRACE: False
  # number of concurrent python processes.
  PARALLELISM: 1

Full example is shown in meshinsight/predictor/config/example1.yml. We provide two sample traces in meshinsight/predictor/CRISP/example1 and meshinsight/predictor/CRISP/example2

Sidecar Proxy Configurations and workload characteristics

If you use the CRISP option, MeshInsight will automatically parse the proxy configurations if you give MeshInsight the deployment files (i.e., Kubernetes yaml files). If MeshInsight cannot match the service name in the trace and in the deployment files, it will assume the proxy is configured as a TCP proxy. In addition, because Jaeger traces does not include workload characteristics, MeshInsight will use a default size (100B) and rate (1000 Req/Sec) for all calls in the call graph. (You can override these values with --size and --rate)

Option 2

This option will be added soon.

Quantifying the impact of service mesh optimizations

MeshInsight an also predict the end-to-end impact of an optimization (e.g., using Unix domain socket instead of loopback). To enable this estimation, you need to provide information (i,e., a speedup profile) on the impact of the optization for the component(s) you have optimized. See meshinsight/predictor/config/speedup.yml for an example speedup profile.

Alternatively, new performance profiles may be based on running offline profiling after implementing the optimization.