Clarify the impact on refreshes for node-stats telemetry

[danielmitterdorfer]

In the docs of the node stats telemetry device we state:

Using this telemetry device will skew your results because the node-stats API triggers additional refreshes. Additionally a lot of metrics get recorded impacting the measurement results even further.

Furthermore, when this telemetry device is used, we issue the following warning:

[WARNING] You have enabled the node-stats telemetry device, but requests to the _nodes/stats Elasticsearch endpoint trigger additional refreshes and WILL SKEW results.

This is only the case before Elasticsearch 7.2 and has been fixed with elastic/elasticsearch#40458. Therefore we should clarify that this only applies to Elasticsearch before 7.2 and maybe issue the warning also only for versions before 7.2.

[bartier]

I'd like to work on this if possible. =)

Can you provide more information how can I get information about Elasticsearch that is going to be used?

Looking at the classes, I found BareProvisioner that I think just provision the distribution that is set in the command line, is there any idea how can I get this information in the NodeStats class? I'm actually newbie at rally yet haha.

[danielmitterdorfer]

I'd like to work on this if possible. =)

Sure, please go ahead. :)

Can you provide more information how can I get information about Elasticsearch that is going to be used?

You can have a look how it is done in ClusterEnvironmentInfo#on_benchmark_start() by calling the node info API. Note that the NodeStats telemetry device can work with multiple clusters (e.g. necessary for cross-cluster-replication) and you need to extract the proper client first, i.e. something along the lines of:

default_client = self.clients["default"]
info = default_client.info()

We have a utility function in esrally.utils.versions that is called components which will split that version string into major, minor, patch etc. which you can use for the version check.

Summarizing the goals of this PR:

• The warning should only be emitted for clusters < version 7.2.0
• The docs should contain a hint that this is only a problem for clusters < version 7.2.0

I hope that helps.