Dgraph: Q: What are my obligations for generating a consistent backup?

[dzrw]

I would like to be able to backup a dgraph cluster hosted on AWS by instructing the cluster to go into a read-only state and flush any write buffers to disk, and then generate EBS snapshots of each disk rather than going through the documented export process. The copy-on-write semantics of EBS snapshots dramatically shorten the maintenance window that I'd need to take to perform a cluster backup over the documented export option.

Now, I realize that dgraph is not an AWS-only product, so I'm not asking anyone for support for my own particular use case.

However, I would like to understand my obligations for generating a consistent backup using the documented export option given this worrying comment.

1) Is there a way to instruct the cluster to reject writes and go into a read-only mode?
2) If not, do I need to gate writes to the cluster through a custom external service?
3) If not, can you clarify how the documented export will generate a consistent backup while concurrently applying mutations?

Thank you; I'm really interesting in dgraph -- it's a good fit for my use case -- but I need clarity on backup handling and disaster recovery; otherwise, I'm going to end up using AWS Neptune (and flirt with writing a GraphQL to SPARQL transpiler which I do not want to do).

[janardhan1993]

@politician Dgraph by design maintains versions of data to support transactions. Dgraph's export reads a snaphost of the database and it's ready only, so you don't need to worry about the ongoing mutations.

[dzrw]

@janardhan1993 Thanks; that makes sense. Does the export filename contain the version number? Generally, given a cluster of 5 nodes, I understand that I have to collect the exported files off of each node and move them to, say, S3 for archival. Later, if I need to restore them, can I restore 1 of 5, or does the tool enforce that all 5 files are present?

[pawanrawal]

• The export file name doesn't contain the version number, though we could make it part of the file name if it's useful.

• There would be an export file on every node which is the leader of a group, so if you had replicas, only one of them would write the exported file.

• You can restore only one if you want to but that would mean you just load partial data for a group.

[manishrjain]

Dgraph supports MVCC, so a given read timestamp would generate a consistent snapshot of the DB. We're soon going to launch backup support to S3 as an enterprise feature, which you can use for this purpose.

[geoyws]

Can we do backups without enterprise features?

[OmarAyo]

Hi @geoyws

Backups are only enabled when a valid license file is supplied to a Zero server OR within the thirty (30) day trial period, no exceptions.