fix(model-server): do not execute long-running repository operation on request threads

[odzhychko]

Execute them on Dispatchers.IO because they are mostly blocked by waiting for data from store. Long-running operations are computeDelta and mergeChanges.

[slisson]

The problematic cases are where a thread is blocked and not doing anything. If there are a lot of request that are blocked at the same location we need a lot of threads.

Using withContext(Dispatchers.IO) is not the best solution, because it just uses the unconstrained thread pool, but this still creates a lot of unnecessary threads. A better solution is to replace the blocking calls by coroutine features.

From the changes in this PR I identified the following problematic calls:

---

The transactions.txStart() call in IgniteStoreClient.runTransaction.

We can wrap that code with a Mutex to avoid trying to start a transaction while there is already an active one. This suspends the coroutine instead of blocking the thread.

---

The usage of lazy in RepositoriesManager.computeDelta.

This can be replaced by a Deferred data type and an async call.

---

Both changes will force you to make the methods, and all the transitive usages, suspendable. This automatically fixes all the locations where you now added withContext(Dispatchers.IO), which is then not required anymore.

I guess this also answers the question

On which level do we want to decide about execution (e.g. which thread pool to execute computation on)

We avoid blocking calls by using coroutine features. If not possible, we use withContext(Dispatchers.IO) around that call. The propagated suspend keyword automatically ensures that all HTTP endpoints are fixed.

[slisson]

🎉 This PR is included in version 5.1.1 🎉

The release is available on GitHub release

Your semantic-release bot 📦🚀