Due to internals on Fastly end, hits are not shared between the various nodes in a POP cluster at fastly. This means that every node located at a POP will fetch from origin even though a different node in the POP cluster already has done so.

The reason for this is that we on our .vcl does a restart, which Fastly incorrectly treats like a error condition.

Fastly support has suggested the following possible solution:

For your use case, clustered behavior may work if you create a request header Fastly-Force-Shield:1 this should take care of making sure it is the primary or secondary that do the requests to origin, and make sure that your objects are cached for the whole cluster.
 
I've also found the similar discussion https://community.fastly.com/t/inconsistent-cache-lookup/305/8 in which I referenced the solution.