[jira] [Created] (SOLR-13234) Prometheus Metric Exporter Not Threadsafe

["Danyal Prout (JIRA)" <ji...@apache.org>]

Danyal Prout created SOLR-13234:
-----------------------------------

             Summary: Prometheus Metric Exporter Not Threadsafe
                 Key: SOLR-13234
                 URL: https://issues.apache.org/jira/browse/SOLR-13234
             Project: Solr
          Issue Type: Bug
      Security Level: Public (Default Security Level. Issues are Public)
          Components: metrics
    Affects Versions: 7.6, 8.0
            Reporter: Danyal Prout
             Fix For: 8.x, master (9.0)


The Solr Prometheus Exporter collects metrics when it receives a HTTP request from Prometheus. Prometheus sends this request, on its [scrape interval|https://prometheus.io/docs/prometheus/latest/configuration/configuration/#scrape_config]. When the time taken to collect the Solr metrics is greater than the scrape interval of the Prometheus server, this results in concurrent metric collection occurring in this [method|https://github.com/apache/lucene-solr/blob/master/solr/contrib/prometheus-exporter/src/java/org/apache/solr/prometheus/collector/SolrCollector.java#L86]. This method doesn’t appear to be thread safe, for instance you could have concurrent modifications of a [map|https://github.com/apache/lucene-solr/blob/master/solr/contrib/prometheus-exporter/src/java/org/apache/solr/prometheus/collector/SolrCollector.java#L119]. After a while the Solr Exporter processes becomes nondeterministic, we've observed NPE and loss of metrics.

To address this, I'm proposing the following fixes:

1. Read/parse the configuration at startup and make it immutable. 
 2. Collect metrics from Solr on an interval which is controlled by the Solr Exporter and cache the metric samples to return during Prometheus scraping. Metric collection can be expensive, for example executing arbitrary Solr searches, it's not ideal to allow for concurrent metric collection and on an interval which is not defined by the Solr Exporter.

There are also a few other performance improvements that we've made while fixing this, for example using the ClusterStateProvider instead of sending multiple HTTP requests to each Solr node to lookup all the cores.

I'm currently finishing up these changes which I'll submit as a PR.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: dev-help@lucene.apache.org