[tarantool-patches] Re: [tarantool-patches] Re: [PATCH v2] replication: do not fetch records twice

Konstantin Belyavskiy k.belyavskiy at tarantool.org
Mon Aug 6 14:03:22 MSK 2018


Hello, Vlad.
Thank you for the review. I fixed issues you had mentioned.
In comments below I will try to explain my position regarding to several design proposal.
Please find an updated RFC in attachment.

>Вторник, 10 июля 2018, 18:16 +03:00 от Vladislav Shpilevoy <v.shpilevoy at tarantool.org>:
>
>Hello. Thanks for the RFC! See my comments below.
>
>> +Building such topology is possible based on following principles:
>> +- Each node is required to notify all his downstream peers (replicas) in case of changes with his upstream subscription configuration. It could be done by add/update/delete records in **_routing** table.
>> +- The connection with lesser count of intermediate nodes has the highest priority. Lets define the number of edges between two peers as a Depth. So if A has direct connection with B, then Depth is 1 and if A connected with C through B, then Depth is 2. So if direct path between two nodes exists then it should be used in downstream peer subscription.
>> +- In case of equal Depth connections first wins. But if shorter path is found, then node first should reconnect and then notify downstream peers with updated paths.
>> +
>> +**_routing** table details.
>> +
>> +| Subscriber | Replication Source | Subscribed via | Depth   |
>> +| :--------- | :----------------- | :------------- | :------ |
>> +| UUID       | UUID               | UUID           | Integer |
>> +
>> +So peer notifies his downstream peers (replicas) with updates in **_cluster** table which replicates to peers. 
>
>Why _cluster? I thought that you use _routing table. And by the way, please, say
>why we can not reuse _cluster for this? To be honest, I very upset at how much
>new system spaces we are creating now: _routing here, _promotion in my patch.
>
>Can't we find a way how to make them optional? How to create such system
>replicaset spaces on demand? 
Using system spaces seems the easiest option since it's guarantees luck of conflicts.
Since it's only a draft proposal, final format of such space is not yet defined. After finishing
this I will look for possibilities to not occupy extra system space slot, but reuse existing one
or create on demand if it's necessary. But right now let me assume that this is a stand alone
table defined within system space.
>
>> When transmitting updated path to next level, increment Depth by one (if B connected to A with
>> depth X and B is has only one upstream C then record will be {UUID_C, UUID_A, UUID_B, X+1}).
>
>Typo 'is has'.
>
>And I can not understand. Please, explain in layman's terms. Here
>C has replication source A and B, right?
>And A has replication source B.
>So C subscribed on B via A?
>
>I would like to see here concrete box.cfg examples for each instance. 
Rewrite this section to make it more clear.
>> +
>> +### List of changes
>> +
>> +1. Extend IPROTO_SUBSCRIBE command with a list of server UUIDs for which SUBSCRIBE should fetch changes. Store this UUIDs within applier's internal data structure. By default issuing SUBSCRIBE with empty list what means no filtering at all.
>> +2. Implement white-list filtering in relay. After processing SUBSCRIBE request, relay has a list of UUIDs. Extract associated peer ids and fill in a filter. By default transmit all records, unless SUBSCRIBE was done with at least one server UUID. In latter case drop all records except originating from replicas in this list.
>> +3. After issuing REQUEST_VOTE to all peers, subscription logic knows a map of server UUIDs, their peers and their vclocks. For each reachable UUID select shortest path and assign UUIDs to direct peer through it this pass goes. Issue the subscribe requests. Notify downstream peers with new topology.
>> +4. Rebalancing. Connect/disconnect should trigger logic to start reassigning process.
>> + - On disconnect first find "orphan" and then reassigned all reachable UUIDs to direct peers through who shortest path goes. Notify downstream peers.
>> + - On connect, by iterating through appliers list, find UUIDs with shorter path found, reassign them to correct peers and issue SUBSCRIBE for recently connected applier and for the one from whom we get these UUIDs back.
>> +
>
>Why is so different connect/disconnect events processing? As I understand, in both cases you should
>recalculate optimal routes in the same way, from scratch.
Updated this section. My motivation to separate connect and disconnect is because
disconnect could cause a race condition if, for example A has a downstreams B and C,
which also connected together. Now imagine situation, that A disconnects and both
B and C try to subscribe to A through each other. In Rationale and alternatives section
I provide more details with two alternative solutions for this possible conflict.
>
>> +### Details and open questions
>> +
>> +On connect (new client or the old one reconnects) two options are available:
>> +1. SUBSCRIBE only to direct peer and wait for updates in **_cluster** to initiate further subscriptions.
>> +2. SUBSCRIBE without any UUIDs (that means subscribe to all).
>> +
>> +## Rationale and alternatives
>> +
>> +### Topology Discovering
>> +
>> +Instead of **_cluster** table updates,
>
>Again: _cluster or _routing? 
_routing, fixed.
>
>> encoded _iproto_ messages could be used. In this case, on every change in peer upstream topology, it should send a Map of *{UUID: depth}* representing its current list of subscriptions to all downstream peers (excluding subset of subscriptions obtaining from this peer in master-master configuration).
>> +
>> +### On network configuration change
>> +
>> +On network configuration change what first, to notify peers or try to resubscribe?
>> +1. If the peer is a direct peer, then we have most recent information about this node based on connection status. If available subscribe and immediately notify downstream peers.
>> +2. On disconnect, it's more complex since in a connected subset if some node is disconnected, others can try to reconnect to this dead node through other nodes, but they do decisions based on old information resulting to massive resubscribe request (node A thinks that it has connection to C through B, but B thinks that it is connected through A). So I think, first need to notify replicas, that connection is dropped, and if other path is available try to resubscribe and then notify all downstream again. Or need to think about some kind of acknowledgement since it could be based on outdated information.
>> +3. On shorter path found, first resubscribe, then notify downstream peers.
>> +4. Balancing. It's possible to slightly extend topology with number of peers subscribed for balancing but does it really needed?
>> 
>


Best regards,
Konstantin Belyavskiy
k.belyavskiy at tarantool.org
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.tarantool.org/pipermail/tarantool-patches/attachments/20180806/ceeb8cca/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: topology_discovering_protocol.md
Type: application/octet-stream
Size: 8052 bytes
Desc: not available
URL: <https://lists.tarantool.org/pipermail/tarantool-patches/attachments/20180806/ceeb8cca/attachment.obj>


More information about the Tarantool-patches mailing list