My Profile Photo

Stefan Prodan


DevOps Consultant. Passionate about software architecture, domain driven design and DevOps. Loves programming in .NET, JS and Go


Docker containers log transport and aggregation at scale

As a DevOps consultant I often deal with large setups, where logging is a critical part of the system. Building a self-hosted centralized logging solution that can handle hundreds of Docker hosts with several million log entries per minute can be a challenging task.

My approach was to use the ELK stack along with a Redis broker. Having a broker between the Docker hosts and the ElasticSearch cluster has many benefits, for once Redis is much faster at data ingestion than ES and you can avoid loosing data if the ES cluster goes offline for a short period of time. Another advantage would be that scaling your system doesn’t relies on ES only, you can increase the ingest volume by adding more Redis and Logstash indexer nodes.

For the Docker hosts I’ve opted for the GELF driver to ship the logs directly to Logstash over UDP. Using this approach means that there are no log files stored on the hosts, this will ease the burden on the hosts disks and speed up the log shipping. If your applications have a high log rate, you should consider pushing the logs directly to Logstash and avoid STDOUT/GELF to limit the CPU usage of the Docker daemon.

For the log parsing you have two options: you can have filters on each host at Logstash Shipper level or if you want the Docker hosts CPU to deal only with apps load, you can chose to run the filtering on the Logstash Indexers nodes.

The logging stack:

  • ElasticSearch Cluster (x3 data nodes)
  • ElasticSearch Ingester (used by Logstash indexer)
  • ElasticSearch Coordinator (used by Kibana)
  • Logstash Indexer
  • Redis Broker (used by Logstash shipper and indexer)
  • Logstash Shipper (used by Docker engine)
  • Kibana
  • Docker GELF driver

Log transport diagram

Flow : container -> docker gelf -> logstash shipper -> redis broker -> logstash indexer -> elasticsearch ingester -> elasticsearch data cluster -> elasticsearch coordinator -> kibana

Flow

Starting with the Redis node and down you can horizontally scale any node. You will need a NGINX reverse proxy if you want to have multiple Kibana nodes.

For testing purposes I’ve build the whole stack with Docker Compose, what follows is a step by step guide of the services composition.

Network setup

I’ve create a dedicated Docker network so each container can have a fix IP.

 docker network create --subnet=192.16.0.0/24 elk

In production you should use an internal DNS server and use domain names instead of fix IP addresses. Each service should reside on a dedicated host. The containers can be started with --net=host to bind directly to the host network to avoiding Docker bridge overhead.

Elasticseach Nodes

All Elasticseach nodes are using the same Docker image that contains the HQ and KOPF management plugins along with a health check command.

Dockerfile

FROM elasticsearch:2.4.3

RUN /usr/share/elasticsearch/bin/plugin install --batch royrusso/elasticsearch-HQ
RUN /usr/share/elasticsearch/bin/plugin install --batch lmenezes/elasticsearch-kopf

COPY docker-healthcheck /usr/local/bin/
RUN chmod +x /usr/local/bin/docker-healthcheck

HEALTHCHECK CMD ["docker-healthcheck"]

COPY config /usr/share/elasticsearch/config

Elasticseach data nodes

The Elasticseach data cluster is made out of 3 nodes. In production you should use a dedicated machine for each node and disable memory swapping.

  elasticsearch-node1:
    build: elasticsearch/
    container_name: elasticsearch-node1
    environment:
      ES_JAVA_OPTS: "-Xmx2g"
      ES_HEAP_SIZE: "1g"
    command: >
        elasticsearch 
        --node.name="node1" 
        --cluster.name="elk" 
        --network.host=0.0.0.0 
        --discovery.zen.ping.multicast.enabled=false 
        --discovery.zen.ping.unicast.hosts="192.16.0.11,192.16.0.12,192.16.0.13" 
        --node.data=true 
        --bootstrap.mlockall=true 
    ports:
      - "9201:9200"
      - "9301:9300"
    networks:
      default:
        ipv4_address: 192.16.0.11
    restart: unless-stopped

  elasticsearch-node2:
    build: elasticsearch/
    container_name: elasticsearch-node2
    environment:
      ES_JAVA_OPTS: "-Xmx2g"
      ES_HEAP_SIZE: "1g"
    command: >
        elasticsearch 
        --node.name="node2" 
        --cluster.name="elk" 
        --network.host=0.0.0.0 
        --discovery.zen.ping.multicast.enabled=false 
        --discovery.zen.ping.unicast.hosts="192.16.0.11,192.16.0.12,192.16.0.13" 
        --node.data=true 
        --bootstrap.mlockall=true 
    ports:
      - "9202:9200"
      - "9302:9300"
    networks:
      default:
        ipv4_address: 192.16.0.12
    restart: unless-stopped

  elasticsearch-node3:
    build: elasticsearch/
    container_name: elasticsearch-node3
    environment:
      ES_JAVA_OPTS: "-Xmx2g"
      ES_HEAP_SIZE: "1g"
    command: >
        elasticsearch 
        --node.name="node3" 
        --cluster.name="elk" 
        --network.host=0.0.0.0 
        --discovery.zen.ping.multicast.enabled=false 
        --discovery.zen.ping.unicast.hosts="192.16.0.11,192.16.0.12,192.16.0.13" 
        --node.data=true 
        --bootstrap.mlockall=true 
    ports:
      - "9203:9200"
      - "9303:9300"
    networks:
      default:
        ipv4_address: 192.16.0.13
    restart: unless-stopped

Elasticseach ingest node

The ingest node acts as a reverse proxy between Logstash Indexer and the Elasticsearch data cluster, when you scale out the data cluster by adding more nodes you don’t have to change the Logstash config.

  elasticsearch-ingester:
    build: elasticsearch/
    container_name: elasticsearch-ingester
    command: >
        elasticsearch 
        --node.name="ingester" 
        --cluster.name="elk" 
        --network.host=0.0.0.0 
        --discovery.zen.ping.multicast.enabled=false 
        --discovery.zen.ping.unicast.hosts="192.16.0.11,192.16.0.12,192.16.0.13" 
        --node.master=false 
        --node.data=false 
        --node.ingest=true 
        --bootstrap.mlockall=true 
    ports:
      - "9221:9200"
      - "9321:9300"
    networks:
      default:
        ipv4_address: 192.16.0.21
    restart: unless-stopped

Elasticseach coordinator node

The coordinator node acts as a router between Kibana and the Elasticsearch data cluster, his main role is to handle the search reduce phase.

  elasticsearch-coordinator:
    build: elasticsearch/
    container_name: elasticsearch-coordinator 
    command: >
        elasticsearch 
        --node.name="coordinator"
        --cluster.name="elk" 
        --network.host=0.0.0.0 
        --discovery.zen.ping.multicast.enabled=false 
        --discovery.zen.ping.unicast.hosts="192.16.0.11,192.16.0.12,192.16.0.13" 
        --node.master=false 
        --node.data=false 
        --node.ingest=false 
        --bootstrap.mlockall=true 
    ports:
      - "9222:9200"
      - "9322:9300"
    networks:
      default:
        ipv4_address: 192.16.0.22

Kibana

The Kibana Docker image contains the yml config file where the Elasticseach coordinator node URL is specified and a startup script that waits for the Elasticseach cluster to be available.

kibana.yml

elasticsearch.url: "http://192.16.0.22:9200"

entrypoint.sh

echo "Waiting for Elasticsearch"
while true; do
    nc -q 1 192.16.0.22 9200 2>/dev/null && break
done

echo "Starting Kibana"
exec kibana

Dockerfile

FROM kibana:4.6.2

RUN apt-get update && apt-get install -y netcat bzip2

COPY config /opt/kibana/config

COPY entrypoint.sh /tmp/entrypoint.sh
RUN chmod +x /tmp/entrypoint.sh

CMD ["/tmp/entrypoint.sh"]

Service definition

  kibana:
    build: kibana/
    container_name: kibana
    ports:
      - "5601:5601"
    restart: unless-stopped

Redis Broker

The Redis Broker acts as a buffer between the Logstash shippers nodes and the Logstash indexer. The more memory you give to this node the longer you can take offline the Logstash indexer and the Elasticseach cluster for upgrade/maintenance work. Shutdown the Logstash indexer and monitor Redis memory usage to determine how log does it take for the memory to fill up. Once the memory fills up Redis will OOM and restart. You can use Redis CLI and run LLEN logstash to determine how many logs your current setup holds.

I’ve disabled Redis disk persistance to max out the insert speed:

redis.conf

#save 900 1
#save 300 10
#save 60 10000

appendonly no

Dockerfile

FROM redis:3.2.6

COPY config /usr/local/etc/redis

Service definition

  redis-broker:
    build: redis-broker/
    container_name: redis-broker
    ports:
      - "6379:6379"
    networks:
      default:
        ipv4_address: 192.16.0.79
    restart: unless-stopped

Logstash Indexer

The Logstash Indexer node pulls the logs from the Redis broker and pushes them into the Elasticseach cluster via the Elasticseach Ingester node.

logstash.config

input {
  redis {
    host => "192.16.0.79"
	port => 6379
    key => "logstash"
    data_type => "list"
    codec => json
  }
}

output {
    elasticsearch {
        hosts => "192.16.0.21:9200"
    }
}

Dockerfile

FROM logstash:2.4.0-1

COPY config /etc/logstash/conf.d

Service definition

  logstash-indexer:
    build: logstash-indexer/
    container_name: logstash-indexer
    command: -f /etc/logstash/conf.d/
    restart: unless-stopped

Logstash Shipper

The Logstash Shipper node listens on UDP and receives the containers logs from Docker engine via the GELF driver. The incoming logs are proccesed using Logstash filers and pushed to the Redis Broker node. You should deploy a Shipper node on each Docker host you want to collect logs from. In production you can expose the UDP port on the host and configure Docker engine GELF driver to use it.

logstash.config

input {
  gelf {
    type => docker
    port => 12201
  }
}

# filters ...

output {
  redis {
    host => "192.16.0.79"
    port => 6379
    data_type => "list"
    key => "logstash"
    codec => json
  }
}

Dockerfile

FROM logstash:2.4.0-1

COPY config /etc/logstash/conf.d

Service definition

  logstash-shipper:
    build: logstash-shipper/
    container_name: logstash-shipper
    command: -f /etc/logstash/conf.d/
    ports:
      - "12201:12201"
      - "12201:12201/udp"
    networks:
      default:
        ipv4_address: 192.16.0.30
    restart: unless-stopped

Collect NGINX logs

I’m going to use a NGINX container to showcase the logs collection. First you need to forward the NGINX logs to STDOUT and STDERR for Docker GELF to capture them.

Dockerfile

FROM nginx:latest

# forward nginx logs to docker log collector
RUN ln -sf /dev/stdout /var/log/nginx/access.log
RUN ln -sf /dev/stderr /var/log/nginx/error.log

Next you need to run NGINX container with the Docker GELF log driver like so:

  nginx:
    build: nginx/
    container_name: nginx
    ports:
      - "80:80"
      - "443:443"
    restart: unless-stopped
    logging:
      driver: gelf
      options:
        gelf-address: "udp://192.16.0.30:12201"
        tag: "nginx"

In production you will probably don’t want to specify the GELF address option for each service. If you expose the Logstash Shipper UDP port on the host you can set the address at the Docker engine level, so all your containers will use GELF as the default logging driver.

Docker Engine GELF config

$ dockerd \
    --log-driver=gelf \
    --log-opt gelf-address=udp://localhost:12201 

You can use the GELF tag option to write custom Logstash filers for each service type. Here is an example of how you can target the NGINX logs inside the Logstash Shipper config:

filter {
  if [tag] == "nginx" {
    # nginx access log 
    if [level] == 6 {
      grok {
        match => [ "message" , "%{COMBINEDAPACHELOG}+%{GREEDYDATA:extra_fields}"]
        overwrite => [ "message" ]
      } 

      date {
        match => [ "timestamp" , "dd/MMM/YYYY:HH:mm:ss Z" ]
        remove_field => [ "timestamp" ]
      }
      
      useragent {
        source => "agent"
      }

      mutate {
        convert => { "response" =>"integer" }
        convert => { "bytes" =>"integer" }
        convert => { "responsetime" =>"float" }
        rename => { "name" => "browser_name" }
        rename => { "major" => "browser_major" }
        rename => { "minor" => "browser_minor" }
      }
    }
    # nginx error log 
    if [level] == 3 {
      grok {
        match => [ "message" , "(?<timestamp>%{YEAR}[./-]%{MONTHNUM}[./-]%{MONTHDAY}[- ]%{TIME}) \[%{LOGLEVEL:severity}\] %{POSINT:pid}#%{NUMBER}: %{GREEDYDATA:errormessage}(?:, client: (?<client>%{IP}|%{HOSTNAME}))(?:, server: %{IPORHOST:server})(?:, request: %{QS:request})?(?:, upstream: \"%{URI:upstream}\")?(?:, host: %{QS:host})?(?:, referrer: \"%{URI:referrer}\")"]
        overwrite => [ "message" ]
      } 

      date {
        match => [ "timestamp" , "YYYY/MM/dd HH:mm:ss" ]
        remove_field => [ "timestamp" ]
      }
    }
  }
}

Note that GELF will set the level field to 6 when the input comes from STDOUT and 3 when it’s STDERR so you can write different filers based on that.

Using the above filter will result in the following Kibana view:

Kibana

The Logstash Grok filter parses the NGINX access log and extracts the http version, http verb, request path, referrer, response http code, bytes, client IP, device, browser and OS details. With these fields you can turn Kibana into a versatile traffic analytics tool.

The GELF driver adds the image name/id, container name/id and the Docker command used to start the container. These fields could prove useful if you’ll want to measure the impact of a new deployment based on logs trending.

Running the stack with Docker Compose

Clone the dockelk repo and run the full stack along with a NGINX container for testing.

$ git clone https://github.com/stefanprodan/dockelk.git
$ cd dockelk
$ bash setup.sh

# wait for Kibana to connect to the cluster
$ docker-compose logs -f kibana

# generate NGINX logs
curl http://localhost
curl http://localhost/404

# open Kibana in browser at http://localhost:5601

If you have any suggestion on improving dockelk please submit an issue or PR on GitHub. Contributions are more than welcome!

comments powered by Disqus