Use a truly random Destination for DestinationAnyShard

[eeSeeGee]

Signed-off-by: Aaron Young [email protected]

[deepthi]

LGTM

[rafael]

This is great! We've noticed in this in the past and wanted to fix it

[rafael]

However, tests are going to fail because they are expecting to be in shard one. One approach that we discussed to address this is that the shard "picker" should be an interface. That way we can use random in the real one, but a mocked one that always picks the first shard that is used in the context of the tests.

[xhh1989]

We JDer use this function for pined table that pined to the first shard.

[eeSeeGee]

We JDer use this function for pined table that pined to the first shard.

That's... weird? Why not use a separate keyspace?

[sougou]

We JDer use this function for pined table that pined to the first shard.

That's... weird? Why not use a separate keyspace?

Do you have this code in your branch: https://github.com/vitessio/vitess/blob/master/go/vt/vtgate/planbuilder/from.go#L193. If so, pinned tables should be routed to the shards they belong to.

[sougou]

@xhh1989 Can you confirm that you have the code I pointed at? If so, this change should not affect you.

[xhh1989]

we have a plan upgrade to version 3.0, there a big difference between JD version and 3.0, we add a single shard table feature based on version 2.1 and that single shard table like pined table but pined on the first shard, use DestinationAnyShard for routing the first shard. @sougou you can merge the PR, I won't let this PR affect our internal version upgrade.