auto-sharding: distributed writes

[githubmanticore]

At present, a distributed table can only manage select and update operations. The goal is to enable it to process insert, replace, and delete operations as well.

Here is a high-level specification of the desired functionality:

When writing to a distributed table with the following configuration: t type='distributed' local='t_2' local='t_1' agent='<node3>:t_3|<node1>:t_3':

• Route each document to its appropriate shard based on its ID
• If a shard is unavailable but has a mirror, use the mirror as a failover
• Clearly indicate in the error message which data was not processed if a shard returns an error while attempting to insert/replace a batch
• Prevent routing write queries to a node that is currently a replication donor to avoid stalls

Before implementing this task, a more comprehensive and detailed specification is necessary. Some questions to consider include:

• Which sharding method should be employed?
• If we have to store a shards map in Manticore, what's the best option for that? Should it be a Manticore RT table, a hidden table? How can a table be hidden in Manticore?
• How can we determine if an agent is a replication donor? Should it fail on that side and then trigger a retry?

It is essential to review and consider the discussions and decisions from https://gitlab.com/manticoresearch/dev/-/issues/1748 before proceeding. https://gitlab.com/manticoresearch/dev/-/issues/2935#note_1151535983 is also worth reading to understand the whole auto-sharding idea.

[githubmanticore]

➤ Sergey Nikolaev commented:

Ilya suggested the idea of implementing the whole auto-sharding functionality in Manticore Buddy with optional implementation of hot places in C++ and using them from PHP. It's to be discussed.

[marclaporte]

I can't access the links from above: https://gitlab.com/manticoresearch/dev/-/issues/1748 https://gitlab.com/manticoresearch/dev/-/issues/2935#note_1151535983