It is not easy to deduce design decisions and reasons from the code, so I would like to see a document that would answer the following questions:

• When and how the schema is reloaded.
• Where it leans on built-in net.box schema reload and where it doesn't.
• Which alternatives were considered and why they were declined?
• The same questions about sharding key and sharding function information reload.

See the following discussion: #111.

See == conditions in schema.batching_result_needs_reload() and sharding.batching_result_needs_sharding_reload() from PR #232.

NB: net.box space objects can update in background, correct? So we can flatten an object using an old schema, receive the new space hash from the storage and compare it with the new schema on the router (because it was updated).