How we can enable IPFS replication?

[aledbf]

First, thank you for the IPFS example, that just works.

Now, I want to replicate the content to other nodes (the autoscaler could remove a node) and I only see IPFS Cluster on top IPFS. Is that correct? Have you tried something similar to that? Adding IPFS cluster suggests statefulset instead of a daemonset, which defeats the purpose of a variable number of nodes.

Thanks in advance for any suggestion/idea

[ktock]

@aledbf Nerdctl + IPFS Cluster on DaemonSet should be possible using something like https://github.com/ktock/stargz-snapshotter/blob/797cf83cabfb6d19240817a665ed6bab78837fc4/script/nerdctl-ipfs-registry-kubernetes-cluster/nerdctl-ipfs-registry.yaml (not well tested. we'll try it deeper and add docs).
Note that 9095 needs to be used as IPFS API port instead of 5001 https://cluster.ipfs.io/documentation/reference/proxy/

[aledbf]

@ktock thank you for the quick response and the ipfs-cluster configuration.

The ipfs cluster configuration is similar to what I have with one small difference, I used replication_factor_max=4.
When I do that, the resource utilization is really different.

Does this make sense to you?

root@server-ws-ephemeral84-m0b3:/home/aledbf# kubectl top pods -n storage
NAME                             CPU(cores)   MEMORY(bytes)   
ipfs-4vnph                       762m         52611Mi         
ipfs-bootstrap-b76c4dc7b-fzc6m   7m           55Mi            
ipfs-dl92k                       21m          1260Mi          
ipfs-fpz7q                       102m         835Mi           
ipfs-hgbb7                       12m          500Mi           
ipfs-jxwxd                       41m          1276Mi          
ipfs-t8xpc                       792m         52521Mi         

the two pods with high memory are being used for pull/push (and running the replication)

[aledbf]

@ktock I find out the reason for the resource utilization issue with ipfs-cluster, it was related to the badgerds profile. After removing that profile the resource utilization problem is gone

Now, the cluster pulling time is measured in milliseconds now.

Again, thank you for the amazing work with nerdctl.