Description

Resolves #4501 and partially resolves #4502

This PR removes --in-memory. Specifically, core will not crash if this flag is passed in for Horizon compatibility reasons, but will ignore it and related flags.

Additionally, --in-memory has been cannibalized to use in testing. Previously, any test that committed changes directly to the root ltx was incompatible with BucketListDB and required SQL. This is an issue, as we are deprecating SQL for ledger state, and it would not make sense to maintain this SQL overhead for testing only. While I've removed most of the in-memory code, I've saved the bits regarding the never committing ledger root to use for the direct ltx commit test cases. Thus we remove the test dependency for SQL.

In order to properly test the orderbook, the in-memory testing mode still maintains and queries a SQL database just for offers (similar to what BucketListDB does).

Originally I did one PR for removing SQL, --in-memory and mandating background eviction. This was way to big to merge but may be helpful in providing context on what changes are in flight after this PR: #4504.

Currently everything works except for the "herder externalizes values" test. With BucketListDB testing mode, nodes can be restarted, but direct ltx commits are not supported. With the in-memory mode testing, ltx can be committed to directly, but nodes can not be restarted.

For all tests except for "herder externalizes values", this if fine. However, "herder externalizes values" is the only test that both directly commits to the ltx and restarts nodes. In particular, the setupUpgradeAtNextLedger function directly commits to the ltx. I've seperated the test into two helper functions, one that doesn't require restart and one that does. I think that the restart test doesn't actually require an upgrade to occur. I think the test just used upgrades in order to populate ledgers with some meaningful, detectable work to make sure externalized ledgers aren't applied several times. If the upgrade operation is not integral to this test, we can probably find a workaround. Alternatively, we can maybe go through the full upgrade arming process with BucketListDB support instead of cutting corners with direct ltx commits. I'm not well versed in herder tests though, so I've left the failing test as is. If we can remove the upgrade and check the condition some other way, this is the path of least resistance.

Test has been fixed to use upgrades compatible with BucketListDB.

Checklist

• Reviewed the contributing document
• Rebased on top of master (no merge commits)
• Ran clang-format v8.0.0 (via make format or the Visual Studio extension)
• Compiles
• Ran all tests
• If change impacts performance, include supporting evidence per the performance document