IMPORTANT: Please read this to learn about the status of this project and alternatives to it.
Based on prometheus-ecs-sd, this project enables your Prometheus to dynamically scrape targets deployed in AWS ECS. The discoverer is perfect if you don't have a service discovery system like Consul in-place. It provides an easy-to-use alternative and can be used in low- to mid-sized environments with hundreds of tasks running in parallel.
Once the discoverer is up and running, any task can be made visible to Prometheus.
What are the advantages of using this project over prometheus-ecs-sd?
- Modified throttling in combination with the already existing caching allows the usage in environments with hundreds of tasks.
- Every target can expose custom labels via its environment variables. This way you can provide more ways to aggregate metrics. For example by type of API.
- Deploy as container (provided and can be used directly) or Python package. No need to manually install boto3 and so on.
- Instrumented with Prometheus. Allows more insights into the discoverer. You can monitor how long discovery rounds take and stuff like used memory. Dashboard already included and ready to use.
- Extensive testing with high coverage ensures functionality.
- More configuration options. For example structured logs.
Still active, but no new features. I recommend using teralytics/prometheus-ecs-discovery.
As of 2021-09 I'm not using PromED anywhere. Everything related to containers is now running on Kubernetes for me. One system (I'm not responsible for anymore) at my job is still using PromED and it has been chucking along just fine for several months now.
Looking at alternative approaches I would now go with AWS Cloud Map instead of PromED with is specific to ECS. It would be possible to refactor PromED to use AWS Cloud Map, but it makes more sense to just recreate a new tool that completely focuses on Cloud Map. Something like this for example:
One big advantage of AWS Cloud Map is that it takes care of providing the correct IP. PromED does this by querying the ECS and EC2 API and has to cover a bunch of corner cases. AWS Cloud Map would simplify this.
As this project is based on prometheus-ecs-sd, the setup is very similar / exactly the same. The setup consists out of three parts. The targets must be prepared by adding environment variables. Next, PromED itself must be deployed. And finally, the Prometheus configuration must be updated.
Targets are setup via setting environment variables in the task definitions.
Set PROMETHEUS_TARGET
to true
to make PromED consider the container. This
by itself is already enough to make it work with the configured defaults.
Alternatively you can use the dockerLabels
in the container definition. For
this you have to set the PromED setting MARKER_TYPE
to dockerLabels
instead of the default environment
.
If your metrics are not exposed on the default /metrics
endpoint, you can
specifiy the endpoint with PROMETHEUS_ENDPOINT
.
You can also declare multiple endpoints and different intervals. The supported
intervals are 15s
, 30s
, 1m
and 5m
. Based on the interval, targets will
end up in different files. The default interval is "generic". Examples for this:
5m:/mymetrics,30s:/mymetrics2
/mymetrics
/mymetrics,30s:/mymetrics2
By default, all targets will end up in a single file called tasks.json
in the
configured directory (defaults to /tmp
).
Sometimes you might want to add additonal labels to targets to group them.
For example by the used API type (REST vs. GraphQL). This can be done by adding
environment variables to the container definition in the respective task
definition with the PROMETHEUS_LABEL_
prefix. For example
PROMETHEUS_LABEL_api_type
or PROMETHEUS_LABEL_foo
. Environment variables
set from within the container are not visible to PromED and are ignored.
"environment": [
{ "name": "PROMETHEUS_LABEL_foo", "value": "bar" },
{ "name": "PROMETHEUS_LABEL_high", "value": "fi" },
]
There is an alternative for this using the dockerLabels
attribute instead.
PromED always checks both locations. Unlike with the environment variables,
the additional target labels are all contained in one string. You can configure
the label PromED should look for by setting CUSTOM_LABELS_KEY
. It works like
this:
"dockerLabels": {
"promed.custom_labels": "foo=bar, high=fi,what=ever"
}
The third option is to use the WITH_DOCKER_LABELS
setting to give PromED an
array of labels it should always try to map from dockerLabels
to target
labels. The PromED config could look like this (notice the config prefix):
"environment": {
"PROMED_WITH_DOCKER_LABELS": "['com.company.namespace', 'com.company.alias']"
}
A container that uses these labels:
"dockerLabels": {
"com.company.namespace": "whatever",
"com-company-alias": "zoomzoom"
}
PromED translates it to com_company_namespace
and com_company_alias
.
Regarding networking, all network modes are supported (bridge
, host
and awsvpc
).
If PROMETHEUS_PORT
and PROMETHEUS_CONTAINER_PORT
are not set, the script
will pick the first port from the container definition (in awsvpc and host
network mode) or the container host network bindings in bridge mode. On
Fargate, if PROMETHEUS_PORT
is not set, it will default to port 80
.
If PROMETHEUS_CONTAINER_PORT
is set, it will look at the container host
network bindings, and find the entry with a matching containerPort
. It will
then use the hostPort
found there as target port. This is useful when the
container port is known, but the hostPort
is randomly picked by ECS (by
setting hostPort
to 0
in the task definition).
If your container uses multiple ports, it's recommended to specify
PROMETHEUS_PORT
(awsvpc
, host
) or PROMETHEUS_CONTAINER_PORT
(bridge
).
Quoted from
prometheus-ecs-sd
commit ece6ca2
.
You can either use the Python package directly or go for the Docker image. Please note that just running the package / image is not enough. The enviornment must be ready for boto3. This includes credentials and approbiate rights.
Available under the name prometheus_ecs_discoverer
. To start PromED:
python -m prometheus_ecs_discoverer.run
To configure PromED you can either provide a settings file or use plain environment variables. Please see Configuration for more info. Please see Configuration for more info.
The image trallnag/prometheus_ecs_discoverer
can be found
here.
The recommended way for configuring the image is to use environment variables.
You will probably want to run the discoverer in ECS. Here, you don't have to provide credentials assuming everything is set up correctly. Boto3 will automatically detect relative credentials URI and retrieve them from AWS. Nevertheless, the region must be set by you.
The actions that PromED is performing on the AWS API can be found in aws-iam-policy.json. The allowed actions must be attached to an approbiate role. If you deploy PromED in ECS, this should look like in aws-iam-ecs-role.json.
If you want all your targets to be scraped in the same interval, the following job is enough.
- job_name: 'ecs'
file_sd_configs:
- files:
- /tmp/tasks.json
relabel_configs:
- source_labels: [metrics_path]
action: replace
target_label: __metrics_path__
regex: (.+)
By default, PromED exposes an /metrics
endpoint.
- job_name: discovery
static_configs:
- targets: ["discovery:8080"]
As this project is based on prometheus-ecs-sd, the configuration is mostly compatible.
PromED uses Dynconf for config management. There
are two main ways you can configure the application. Either by providing a
custom settings file with the env var SETTINGS_FILES_FOR_DYNACONF
(see
here)
or directly setting the respective values via env vars with the PROMED_
prefix. All supported settings together with their default values can be found
settings.toml
(click me).
You can find the Grafana dashboards for PromED in the documents folder.
A second dashboard contains panels with Grafana Alerts enabled.