Logging in Kubernetes: From container to visualization
In our previous blog, we took a look at some basics of logging in kubernetes and how we can use Fluentd and Stakater Konfigurator for dynamic log parsing. In this blog let’s take a look at the processes and tools to build a broader logging ecosystem for our kubernetes cluster.
Log file generation
Let’s first take a look at how a log file is generated on the cluster. All applications are run in docker containers, so how they generate logs is important. At Stakater we follow the best practice of writing our logs to STDOUT stream in the container. This is compliant with the twelve-factor app principle in that, the application running within the container should not concern itself with how to write, buffer or store the logs. It should simply write its stream of logs to STDOUT. It should be the responsibility of the execution environment to capture, process and route the logs in whichever way is suitable. Also, we do not need to separately add volume mounts to map our log files outside of the container to have them be processed, which would add the path of the log file into the equation. Should there be any change in the file path in future, you will need to handle that in volume mount and maybe even any downstream system. Another benefit of writing to STDOUT is that the logs are automatically captured by docker logs, which is not the case for logs that are not written to STDOUT, such as to a custom log file. Docker of course captures these logs based on the logging driver that it is configured with. Using the default
json driver means that the logs are stored in json format on disk in a specific file path.
In case you are using a third party application or tool which writes its logs to a log file rather than STDOUT, you need not worry. There is always a workaround, and one suitable workaround for this is adding a symbolic link (symlink) from the log file to /dev/stdout. An example of this is Nginx, where the server itself writes its logs to a logfile, but the officially provided dockerfile adds the symlink.
Outside of the containers, Kubernetes creates symlinks to the docker log files at /var/log/containers/*.log on the Node. We can therefore easily locate and capture logs within our fluentd daemonset.
We do not handle persistent storage of log files at this point since we have our fluentd daemonset forwarding the logs in real-time. Persistence of the log files will handled at a later point in the logging flow as we will see soon.
Fluentd is an open source data collector for a unified logging layer. It has a flexible plugin architecture, allowing easy extension of its functionality. We deploy it as a DaemonSet on our cluster to ensure that all Nodes run a copy of the fluentd Pod. Any new node that is added to the cluster will also get a fluentd pod automatically.
The fluentd input plugin has the responsibility for reading in data from these log sources, and generating a Fluentd event against it. We use the
in_tail Input plugin which allows Fluentd to read events from the tail of text files. This position from where fluentd has read a particular log file is recorded in a position file. And next time fluentd will pick up reading from this position in the file. The input is parsed, based on the configuration provided to the input plugin.
The fluentd event contains information such as where an event comes from, the time of the event, and the actual log content. We use the
kubernetes_metadata_filter plugin to enrich the log event with kubernetes pod and namespace metadata. The plugin gets basic metadata about the container that emitted a given log record using the source of the log record.
These log events are matched to an Output plugin type in the Fluentd configuration. We use the
out_elasticsearch Output plugin to forward these records to Elasticsearch.
Log rotation essentially renames the old log file, and creates a new log file in its place for continued capturing of logs. This is done so that the file size remains manageable. At Stakater we use the linux logrotate utility for this, also as a daemonset so that it runs on every node like fluentd. At the moment there isn’t a public chart available in the official helm repository, we are using our own which you can find here.
We can see an example configuration as follows. The schedule is maintained through cron and the rest of the parameters such as path and size of file, number of rotations, etc. are specified in the configuration:
cronSchedule: 0 */12 * * *
create 0644 root root
Elasticsearch is basically a NoSQL database that became popular in the Logging domain as part of the ELK/ElasticStack. As we saw before, Fluentd forwards the log streams to Elasticsearch, which then goes on to index and store the log information. Elasticsearch basically allows us to easily search logs based on the metadata that was added by Fluentd. We can easily get results from its REST API to look at logs of a particular application, or apps on a particular node, or even search for all ERROR logs that may have been generated throughout the cluster. A RESTful api however is not a user friendly way of viewing log information, and can also be restricting to one query at a time. A more powerful use of this is to be able to see the log data in terms of trends, and also aggregations. This is where Kibana comes in as we will see in a subsequent section.
With Elasticsearch we would like to have persistent storage, so that our log data and indices are safe in case of a server crash. We there back the Elasticsearch deployment with a Persistent Volume Claim (PVC) on Kubernetes. This way the logs are persisted across Node restarts.
We may also like to archive our Elasticsearch indices instead of having them all maintained since the beginning of time in our EBS. Old logs may not be needed and we can easily move them to a cheaper form of storage, such as an AWS S3 bucket with glacial storage. In such cases we use ES Curator. It helps us to curate our Elasticsearch indices. Following is an example configuration of how the curator cron schedule and other properties can be configured.
cronSchedule: 0 0 * * *
Cerebro is a tool we use for web administration of Elasticsearch. It can be useful to get a web view of the elasticsearch nodes, indices, shards, resource usage etc.
Kibana, is a visualization tool which helps us to view queries on Elasticsearch. We can view the text log data as-is, filter and query it based on certain labels to view just a subset, and we can also view it in the form of a chart, graph or other visualization. Kibana comes with some visualizations out of the box, and we can also build custom visualizations and dashboards as per our requirement. Visualizations can be helpful in indicating trends such as number of warning or error logs throughout the day time, which could be correlated to changing server load or other factors. Such analyses can then help identify issues pro-actively.