Prometheus 2.0 migration guide

In line with our stability promise, the Prometheus 2.0 release contains a number of backwards incompatible changes. This document offers guidance on migrating from Prometheus 1.8 to Prometheus 2.0.

Flags

The format of the Prometheus command line flags has changed. Instead of a single dash, all flags now use a double dash. Common flags (--config.file, --web.listen-address and --web.external-url) are still the same but beyond that, almost all the storage-related flags have been removed.

Some notable flags which have been removed:

  • -alertmanager.url In Prometheus 2.0, the command line flags for configuring a static Alertmanager URL have been removed. Alertmanager must now be discovered via service discovery, see Alertmanager service discovery.

  • -log.format In Prometheus 2.0 logs can only be streamed to standard error.

  • -query.staleness-delta has been renamed to --query.lookback-delta; Prometheus 2.0 introduces a new mechanism for handling staleness, see staleness.

  • -storage.local.* Prometheus 2.0 introduces a new storage engine, as such all flags relating to the old engine have been removed. For information on the new engine, see Storage.

  • -storage.remote.* Prometheus 2.0 has removed the already deprecated remote storage flags, and will fail to start if they are supplied. To write to InfluxDB, Graphite, or OpenTSDB use the relevant storage adapter.

Alertmanager service discovery

Alertmanager service discovery was introduced in Prometheus 1.4, allowing Prometheus to dynamically discover Alertmanager replicas using the same mechanism as scrape targets. In Prometheus 2.0, the command line flags for static Alertmanager config have been removed, so the following command line flag:

  1. ./prometheus -alertmanager.url=http://alertmanager:9093/

Would be replaced with the following in the prometheus.yml config file:

  1. alerting:
  2. alertmanagers:
  3. - static_configs:
  4. - targets:
  5. - alertmanager:9093

You can also use all the usual Prometheus service discovery integrations and relabeling in your Alertmanager configuration. This snippet instructs Prometheus to search for Kubernetes pods, in the default namespace, with the label name: alertmanager and with a non-empty port.

  1. alerting:
  2. alertmanagers:
  3. - kubernetes_sd_configs:
  4. - role: pod
  5. tls_config:
  6. ca_file: /var/run/secrets/kubernetes.io/serviceaccount/ca.crt
  7. bearer_token_file: /var/run/secrets/kubernetes.io/serviceaccount/token
  8. relabel_configs:
  9. - source_labels: [__meta_kubernetes_pod_label_name]
  10. regex: alertmanager
  11. action: keep
  12. - source_labels: [__meta_kubernetes_namespace]
  13. regex: default
  14. action: keep
  15. - source_labels: [__meta_kubernetes_pod_container_port_number]
  16. regex:
  17. action: drop

Recording rules and alerts

The format for configuring alerting and recording rules has been changed to YAML. An example of a recording rule and alert in the old format:

  1. job:request_duration_seconds:histogram_quantile99 =
  2. histogram_quantile(0.99, sum by (le, job) (rate(request_duration_seconds_bucket[1m])))
  3. ALERT FrontendRequestLatency
  4. IF job:request_duration_seconds:histogram_quantile99{job="frontend"} > 0.1
  5. FOR 5m
  6. ANNOTATIONS {
  7. summary = "High frontend request latency",
  8. }

Would look like this:

  1. groups:
  2. - name: example.rules
  3. rules:
  4. - record: job:request_duration_seconds:histogram_quantile99
  5. expr: histogram_quantile(0.99, sum by (le, job) (rate(request_duration_seconds_bucket[1m])))
  6. - alert: FrontendRequestLatency
  7. expr: job:request_duration_seconds:histogram_quantile99{job="frontend"} > 0.1
  8. for: 5m
  9. annotations:
  10. summary: High frontend request latency

To help with the change, the promtool tool has a mode to automate the rules conversion. Given a .rules file, it will output a .rules.yml file in the new format. For example:

  1. $ promtool update rules example.rules

Note that you will need to use promtool from 2.0, not 1.8.

Storage

The data format in Prometheus 2.0 has completely changed and is not backwards compatible with 1.8. To retain access to your historic monitoring data we recommend you run a non-scraping Prometheus instance running at least version 1.8.1 in parallel with your Prometheus 2.0 instance, and have the new server read existing data from the old one via the remote read protocol.

Your Prometheus 1.8 instance should be started with the following flags and an config file containing only the external_labels setting (if any):

  1. $ ./prometheus-1.8.1.linux-amd64/prometheus -web.listen-address ":9094" -config.file old.yml

Prometheus 2.0 can then be started (on the same machine) with the following flags:

  1. $ ./prometheus-2.0.0.linux-amd64/prometheus --config.file prometheus.yml

Where prometheus.yml contains in addition to your full existing configuration, the stanza:

  1. remote_read:
  2. - url: "http://localhost:9094/api/v1/read"

PromQL

The following features have been removed from PromQL:

  • drop_common_labels function - the without aggregation modifier should be used instead.
  • keep_common aggregation modifier - the by modifier should be used instead.
  • count_scalar function - use cases are better handled by absent() or correct propagation of labels in operations.

See issue #3060 for more details.

Miscellaneous

Prometheus non-root user

The Prometheus Docker image is now built to run Prometheus as a non-root user. If you want the Prometheus UI/API to listen on a low port number (say, port 80), you’ll need to override it. For Kubernetes, you would use the following YAML:

  1. apiVersion: v1
  2. kind: Pod
  3. metadata:
  4. name: security-context-demo-2
  5. spec:
  6. securityContext:
  7. runAsUser: 0
  8. ...

See Configure a Security Context for a Pod or Container for more details.

If you’re using Docker, then the following snippet would be used:

  1. docker run -u root -p 80:80 prom/prometheus:v2.0.0-rc.2 --web.listen-address :80

Prometheus lifecycle

If you use the Prometheus /-/reload HTTP endpoint to automatically reload your Prometheus config when it changes, these endpoints are disabled by default for security reasons in Prometheus 2.0. To enable them, set the --web.enable-lifecycle flag.