Caveats¤
Filesystem in Volumes¤
Info
Please avoid to deploy on NFS (AWS EFS) volumes, use block storage instead (e.g. AWS EBS).
In Kubernetes deployments in AWS we saw issues with volumes based on NFS, such as Amazon EFS (Elastic File System).
For DataIntegration Python Plugins we noticed issues: see Sharing the PYTHONPATH via NFS
For GraphDB we saw issues when trying to recover from a previous backups. This is because NFS behaves different why trying to delete files. In Fact GraphDB requires to store the data on EBS (Elastic Block Storage) volumes. See GraphDB technical-requirements.
Load Balancer¤
Depending on the deployment, load balancers or proxies in between Corporate Memory components (such as DataIntegration - DataPlatform, DataPlatform - Database Triplestore) tend to drop long running TCP connections when they are idle. As e.g. SPARQL Update requests may be idle for a long time (while the update is performed) the client will not receive the response. Hence, DI workflow execution will not proceed even though the update went through.
Similar problems can occur between the connection between DataPlatform and GraphDB.
In addition to the hints in the next sections you can always change the TCP keep-alive of the system hosting the containers.
However in cloud environments this often isn’t practical. In Debian based Linux distribution you have to edit /etc/sysctl.conf.
But be advised, only do this, if you are aware of the risks.
net.ipv4.tcp_keepalive_time = 600
net.ipv4.tcp_keepalive_intvl = 60
net.ipv4.tcp_keepalive_probes = 20
Kubernetes Deployments¤
In Kubernetes deployments services for each deployment usually use load balancers to connect other services. However we never saw problems as long as inter-component-communication stays at service level.
In detail, DataIntegration can be set like this in the cmem-helm value file: .Values.dataintegration.config.dataplatformUrl or let it defaulting to: http://dataplatform.namespace:8080/dataplatform.
To connect to Ontotext GraphDB we emphasize to NOT use external load balancers trough an ingress or similar.
Instead use the internal connection like graphdb-node.graphdb.cluster.local at the value .Values.dataplatform.store.graphdb.host.
Container Deployments or distributing over multiple VMs¤
When deploying Corporate Memory in cloud environments such as EC2 where each component runs on a separate VM, or using Container Services such as ECS (Amazon Elastic Container Service) or ACI (Azure Container Instances) the same rules apply as in Kubernetes deployments. For networking make sure to create a Private Network (VPC) instead of bridges to bypass load balancers. VPN
DataIntegration is configured with the ENV variable DATAPLATFORM_URL to set up the connection to DataPlatform or in its config file dataintegration.conf.
For DataPlatform you have to set this in application.yaml or through ENV variables, such as: STORE_GRAPHDB_HOST, STORE_GRAPHDB_PORT and STORE_GRAPHDB_SSL_ENABLED.