Running multiple Skyline instances
It is possible to run Skyline in a somewhat distributed manner in terms of running multiple Skyline instances and those Skyline instances sharing a MySQL database but analysing different metrics and querying the Redis remotely where required, e.g. Luminosity.
Some organisations have multiple Graphite instances for sharding, geographic or latency reasons. In such a case it is possible that each Graphite instance would pickle to a Skyline instance and for there to be multiple Skyline instances. Or you might want to run mulitple Skyline instances due to the number of metrics being handled. However…
Running Skyline in a distributed, HA manner is mostly related to running the components of Skyline in this fashion, for example using Galera cluster to provide MariaDB replication. Each of the components have their own high availability and clustering methods in most cases and the addition of haproxy, mysqlproxy, load balanced or round robin DNS can achieve a lot of redundancy. Defining how Skyline can be run in HA or clustered is beyond the scope of Skyline itself, it is more an exercise of operations and distributed systems which is beyond the scope of this documentation.
However one word of caution, do not cluster Redis. Although some sharded configuration may work, it is simpler just to use local Redis data. Due to the metric time series constantly changing in Redis clustering or slaving results in mostly the entire Redis data store being shipped constantly, not desired.
That said the actual Skyline modules have certain settings and configurations that are HA or distributed _aware_ to allow Skyline itself to provide the ability to use HA/clustered/distributed components. This section deals with these.
The following settings pertain to running multiple Skyline instances:
With the introduction of Luminosity a requirement for Skyline to pull the time series data from remote Skyline instances was added to allow for cross correlation of all the metrics in the population. Skyline does this via a api endpoint and preprocesses the time series data for Luminosity on the remote Skyline instance and returns only the fragments (gzipped) of time series required for analysis, by default the previous 12 minutes, to minimise bandwidth and ensure performance is maintained.
Running Skyline in any form of clustered configuration requires that each Skyline instance know about the other instances and has access to them via the webapp therefore appropriate firewall, network rules and reverse proxy configuration (Apache or nginx) needs to allow this.
settings.REMOTE_SKYLINE_INSTANCES is a required list of alternative URLs,
username, password and hostname for the other instances in the Skyline cluster
so that if a request is made to the Skyline webapp for a resource it does not
have, it can return the other URLs to the client. This is also used in
settings.SYNC_CLUSTER_FILES so that each instance in the cluster can
sync relevant Ionosphere training data and features profiles data to itself.
It is used by Skyline internally to request resources from other Skyline instances to:
Retrieve time series data and general data for metrics served by the other Skyline instance/s.
To retrieve resources for certain client and API requests to respond with all the data for the cluster, in terms of unique_metrics, alerting_metrics, etc.
To sync Ionosphere data between the cluster instances.
Hot standby configuration
Although there are many possible methods and configurations to ensure that single points of failure are mitigated in infrastructure, this can be difficult to achieve with both Graphite and Skyline. This is due to the nature and volume of the data being dealt with, especially if you are interested in ensuring that you have redundant storage for disaster recovery.
With Graphite it is difficult to ensure the whisper data is redundant, due to volume and real time nature of the whisper data files.
With Skyline it is difficult to ensure the real time Redis data is redundant given that you DO NOT want to run a Redis slave as the ENTIRE key store constantly changes. Slaving a Skyline Redis instance is not an option as it will use mountains of network bandwidth and just would not work.
One possible configuration to achieve redundancy of Graphite and Skyline data is to run a Graphite and a Skyline instance as hot standbys in a different data center. Where the primary Graphite is pickling to a primary Skyline instance and a standby Graphite instance. With the standby Graphite instance pickling data to the standby Skyline instance.
| | |
carbon-cache pickle pickle
| +-->--> data-center-2
In terms of the Skyline configuration of the hot standby you configure skyline-2
the same as skyline-1 in terms of alerts, etc, but you set
In the event of a failure of graphite-1 you reconfigure your things to send
their metrics to graphite-2 and set skyline-2
settings.LUMINOSITY_ENABLED to True.
In the event of a failure of skyline-1 you set skyline-2
The setting up of a hot standby Graphite instance requires pickling AND periodic flock rsyncing of all the whisper files from graphite-1 to graphite-2 to ensure that any data that graphite-2 may have been lost in any fullQueueDrops experienced with the pickle from graphite-1 to graphite-2 due to network partitioning, etc, are updated. flock rsyncing all the whisper files daily mostly handles this and ensures that you have no gaps in the whisper data on your backup Graphite instance.
With the addition of labelled_metrics, one can use a VictoriaMetrics cluster to achieve HA of the TSDB data for labelled_metrics.
In terms of the functionality in webapp, the webapp is multiple instance aware. Where any “not in Redis” UI errors are found, webapp responds to the request with a 302 redirect to the remote Skyline instance that is assigned the metric.
Cluster nodes will sync training data and features profiles data between themselves, however currently saved training is not synced between cluster nodes any user saved training data will only be available on the cluster node on which it was saved. Therefore each cluster node has its own saved training data pages. This only relates to training data that is specifically saved by the user and not normal operational training data that is generated for Ionosphere.