prometheus/discovery
Pedro Araújo 575f665944 Add OS type meta label to Azure SD (#3863)
There is currently no way to differentiate Windows instances from Linux
ones. This is needed when you have a mix of node_exporters /
wmi_exporters for OS-level metrics and you want to have them in separate
scrape jobs.

This change allows you to do just that. Example:

```
  - job_name: 'node'
    azure_sd_configs:
      - <azure_sd_config>
    relabel_configs:
      - source_labels: [__meta_azure_machine_os_type]
        regex: Linux
        action: keep
```

The way the vendor'd AzureSDK provides to get the OsType is a bit
awkward - as far as I can tell, this information can only be gotten from
the startup disk. Newer versions of the SDK appear to improve this a
bit (by having OS information in the InstanceView), but the current way
still works.
2018-02-19 15:40:57 +00:00
..
azure Add OS type meta label to Azure SD (#3863) 2018-02-19 15:40:57 +00:00
config Refactor SD configuration to remove `config` dependency (#3629) 2017-12-29 21:01:34 +01:00
consul Set consul server default to localhost:8500 (#3703) 2018-01-24 12:14:32 +00:00
dns comments with TargetProvider should read Discoverer instead (#3667) 2018-01-08 23:59:18 +00:00
ec2 Use shared types from the `common` repo (#3674) 2018-01-11 16:10:25 +01:00
file rename package retrieve to scrape 2018-02-01 09:55:07 +00:00
gce comments with TargetProvider should read Discoverer instead (#3667) 2018-01-08 23:59:18 +00:00
kubernetes Send update when pod's IP address is empty 2018-02-14 14:23:52 +01:00
marathon Use shared types from the `common` repo (#3674) 2018-01-11 16:10:25 +01:00
openstack Use shared types from the `common` repo (#3674) 2018-01-11 16:10:25 +01:00
targetgroup Refactor SD configuration to remove `config` dependency (#3629) 2017-12-29 21:01:34 +01:00
triton Use shared types from the `common` repo (#3674) 2018-01-11 16:10:25 +01:00
zookeeper comments with TargetProvider should read Discoverer instead (#3667) 2018-01-08 23:59:18 +00:00
README.md docs: SD implementations do not have to only send new/changed target groups (#3713) 2018-01-26 22:03:11 +00:00
manager.go remove dicovery race for the context field 2018-01-29 15:18:07 +00:00
manager_test.go updated the sd tests 2018-01-29 15:19:15 +00:00

README.md

Service Discovery

This directory contains the service discovery (SD) component of Prometheus.

There is currently a moratorium on new service discovery mechanisms being added to Prometheus due to a lack of developer capacity. In the meantime file_sd remains available.

Design of a Prometheus SD

There are many requests to add new SDs to Prometheus, this section looks at what makes a good SD and covers some of the common implementation issues.

Does this make sense as an SD?

The first question to be asked is does it make sense to add this particular SD? An SD mechanism should be reasonably well established, and at a minimum in use across multiple organisations. It should allow discovering of machines and/or services running somewhere. When exactly an SD is popular enough to justify being added to Prometheus natively is an open question.

It should not be a brand new SD mechanism, or a variant of an established mechanism. We want to integrate Prometheus with the SD that's already there in your infrastructure, not invent yet more ways to do service discovery. We also do not add mechanisms to work around users lacking service discovery and/or configuration management infrastructure.

SDs that merely discover other applications running the same software (e.g. talk to one Kafka or Cassandra server to find the others) are not service discovery. In that case the SD you should be looking at is whatever decides that a machine is going to be a Kafka server, likely a machine database or configuration management system.

If something is particularly custom or unusual, file_sd is the generic mechanism provided for users to hook in. Generally with Prometheus we offer a single generic mechanism for things with infinite variations, rather than trying to support everything natively (see also, alertmanager webhook, remote read, remote write, node exporter textfile collector). For example anything that would involve talking to a relational database should use file_sd instead.

For configuration management systems like Chef, while they do have a database/API that'd in principle make sense to talk to for service discovery, the idiomatic approach is to use Chef's templating facilities to write out a file for use with file_sd.

Mapping from SD to Prometheus

The general principle with SD is to extract all the potentially useful information we can out of the SD, and let the user choose what they need of it using relabelling. This information is generally termed metadata.

Metadata is exposed as a set of key/value pairs (labels) per target. The keys are prefixed with __meta_<sdname>_<key>, and there should also be an __address__ label with the host:port of the target (preferably an IP address to avoid DNS lookups). No other labelnames should be exposed.

It is very common for initial pull requests for new SDs to include hardcoded assumptions that make sense for the the author's setup. SD should be generic, any customisation should be handled via relabelling. There should be basically no business logic, filtering, or transformations of the data from the SD beyond that which is needed to fit it into the metadata data model.

Arrays (e.g. a list of tags) should be converted to a single label with the array values joined with a comma. Also prefix and suffix the value with a comma. So for example the array [a, b, c] would become ,a,b,c,. As relabelling regexes are fully anchored, this makes it easier to write correct regexes against (.*,a,.* works no matter where a appears in the list). The canonical example of this is __meta_consul_tags.

Maps, hashes and other forms of key/value pairs should be all prefixed and exposed as labels. For example for EC2 tags, there would be __meta_ec2_tag_Description=mydescription for the Description tag. Labelnames may only contain [_a-zA-Z0-9], sanitize by replacing with underscores as needed.

For targets with multiple potential ports, you can a) expose them as a list, b) if they're named expose them as a map or c) expose them each as their own target. Kubernetes SD takes the target per port approach. a) and b) can be combined.

For machine-like SDs (OpenStack, EC2, Kubernetes to some extent) there may be multiple network interfaces for a target. Thus far reporting the details of only the first/primary network interface has sufficed.

Other implementation considerations

SDs are intended to dump all possible targets. For example the optional use of EC2 service discovery would be to take the entire region's worth of EC2 instances it provides and do everything needed in one scrape_config. For large deployments where you are only interested in a small proportion of the returned targets, this may cause performance issues. If this occurs it is acceptable to also offer filtering via whatever mechanisms the SD exposes. For EC2 that would be the Filter option on DescribeInstances. Keep in mind that this is a performance optimisation, it should be possible to do the same filtering using relabelling alone. As with SD generally, we do not invent new ways to filter targets (that is what relabelling is for), merely offer up whatever functionality the SD itself offers.

It is a general rule with Prometheus that all configuration comes from the configuration file. While the libraries you use to talk to the SD may also offer other mechanisms for providing configuration/authentication under the covers (EC2's use of environment variables being a prime example), using your SD mechanism should not require this. Put another way, your SD implementation should not read environment variables or files to obtain configuration.

Some SD mechanisms have rate limits that make them challenging to use. As an example we have unfortunately had to reject Amazon ECS service discovery due to the rate limits being so low that it would not be usable for anything beyond small setups.

If a system offers multiple distinct types of SD, select which is in use with a configuration option rather than returning them all from one mega SD that requires relabelling to select just the one you want. So far we have only seen this with Kubernetes. When a single SD with a selector vs. multiple distinct SDs makes sense is an open question.

If there is a failure while processing talking to the SD, abort rather than returning partial data. It is better to work from stale targets than partial or incorrect metadata.

The information obtained from service discovery is not considered sensitive security wise. Do not return secrets in metadata, anyone with access to the Prometheus server will be able to see them.

Writing an SD mechanism

The SD interface

A Service Discovery (SD) mechanism has to discover targets and provide them to Prometheus. We expect similar targets to be grouped together, in the form of a TargetGroup. The SD mechanism sends the targets down to prometheus as list of TargetGroups.

An SD mechanism has to implement the Discoverer Interface:

type Discoverer interface {
	Run(ctx context.Context, up chan<- []*config.TargetGroup)
}

Prometheus will call the Run() method on a provider to initialise the discovery mechanism. The mechanism will then send all the TargetGroups into the channel. Now the mechanism will watch for changes. For each update it can send all TargetGroups, or only changed and new TargetGroups, down the channel. Manager will handle both cases.

For example if we had a discovery mechanism and it retrieves the following groups:

[]config.TargetGroup{
  {
    Targets: []model.LabelSet{
       {
          "__instance__": "10.11.150.1:7870",
          "hostname": "demo-target-1",
          "test": "simple-test",
       },
       {
          "__instance__": "10.11.150.4:7870",
          "hostname": "demo-target-2",
          "test": "simple-test",
       },
    },
    Labels: map[LabelName][LabelValue] {
      "job": "mysql",
    },
    "Source": "file1", 
  },
  {
    Targets: []model.LabelSet{
       {
          "__instance__": "10.11.122.11:6001",
          "hostname": "demo-postgres-1",
          "test": "simple-test",
       },
       {
          "__instance__": "10.11.122.15:6001",
          "hostname": "demo-postgres-2",
          "test": "simple-test",
       },
    },
    Labels: map[LabelName][LabelValue] {
      "job": "postgres",
    },
    "Source": "file2", 
  },
}

Here there are two TargetGroups one group with source file1 and another with file2. The grouping is implementation specific and could even be one target per group. But, one has to make sure every target group sent by an SD instance should have a Source which is unique across all the TargetGroups of that SD instance.

In this case, both the TargetGroups are sent down the channel the first time Run() is called. Now, for an update, we need to send the whole changed TargetGroup down the channel. i.e, if the target with hostname: demo-postgres-2 goes away, we send:

&config.TargetGroup{
  Targets: []model.LabelSet{
     {
        "__instance__": "10.11.122.11:6001",
        "hostname": "demo-postgres-1",
        "test": "simple-test",
     },
  },
  Labels: map[LabelName][LabelValue] {
    "job": "postgres",
  },
  "Source": "file2", 
}

down the channel.

If all the targets in a group go away, we need to send the target groups with empty Targets down the channel. i.e, if all targets with job: postgres go away, we send:

&config.TargetGroup{
  Targets: nil,
  "Source": "file2", 
}

down the channel.