Platform components
These run once per installation. The chart sets these requests and limits invalues.yaml:
KubeBlocks, the snapshot controller and FalkorDB Browser also run once per installation. The chart sets no requests for them, so they use their upstream charts’ defaults. Override them under
kubeblocks, snapshot-controller and falkordb-browser in the values.
The cluster baseline in Requirements, 3 worker nodes with 4 CPU and 16 GB of memory each, covers the platform components and a small set of deployments.
Deployments by topology
The Admin UI wizard starts every data node at1 CPU, 2Gi of memory and a 10Gi volume, and every Sentinel at 500m CPU, 512Mi and a 5Gi volume. With those defaults, each topology needs:
The wizard’s minimum for a sharded deployment is 3 shards of 2 nodes each. Each database pod also runs the
kbagent and metrics sidecars, which add a small amount on top of the node’s own resources.
The license counts the CPU and memory requests of every node, including Sentinels and every shard’s replicas, falling back to limits when no request is set. Use the totals above when you check a deployment against your license. See Licensing.
Choosing memory
FalkorDB keeps graphs in memory, so memory is the figure that matters most.- Start from the data. Give each data node enough memory for the graphs it holds, with headroom for queries, replication buffers and background saves. On a sharded deployment, a graph is one key and lives on one shard, so the largest graph must fit on one node.
- Replicas need the same memory as the primary. Every replica holds a full copy of its primary’s data.
- Size storage from memory. The data volume holds the persistence files, so make it larger than the memory limit. Volumes can grow later; see Parameters & resources.
- Let memory grow automatically. Automatic memory scaling raises a deployment’s memory as it fills, up to a ceiling you set. See Scale memory automatically.
Choosing CPU
Start at the wizard’s1 CPU per data node. Add CPU when the pods’ CPU stays busy or query latency rises as more queries run at the same time. Monitoring shows both.