This commit builds on the ideas of @cfromknecht in lnd/5153. The
addition is that the design is now simpler and more robust by queueing
up everything, but allowing maximal parallelism where txns don't block.
Furthermore the commit makes CommitQueue.Done() private essentially
removing the need to understand the queue externally.
This commit fixes an issue where subsequent transaction retries may have
changed the read/write sets inside the STM which in turn left junk
references to these keys in the CommitQueue. The left keys potentially
conflicted with subsequent transactions, queueing them up causing
througput degradation.
Depends on btcsuite/btcwallet#757.
Pulls in the updated version of btcwallet and walletdb that have the DB
interface enhanced by their own View() and Update() methods with the
reset callback/closure supported out of the box. That way the global
package-level View() and Update() functions now become pure redirects.
Since we're now storing the content of multiple previously distinct
database files in etcd, we want to properly namespace them as not to
provoke any key collisions. We append a sub namespace to the given
global namespace in order to still support multiple lnd nodes using the
same etcd database simultaneously.
Because the btcwallet code uses the legacy walletdb interface we must
assume it is not fully concurrency safe. Therefore we make sure only a
single writer can be active at any given time for the wallet DB backend
when using etcd.
Since bbolt uses syscalls for memory mapping that aren't available in
JS/WASM builds, we need to make sure we don't reference that code at
all. Otherwise we can't use parts of lnd as a library in projects that
are being compiled down to a WASM binary.
We only want to register the bbolt DB backend ("bdb") when we're not
compiling for a JS/WASM build targets. That's why we want to have that
import in only one file that we can add a build tag to. We remove it in
two other places since only one import is enough anyway.
This commit removes an assertion which is not needed because with etcd
we can safely create keys and values with the same key since they are
stored under different keys in the DB. This saves us one unnecessary Get
on every Put.
The etcd cursor Delete stepped to the next item in the range before
Delete to not invalidate the iteation. This is unnecessary and not
compatible with bbolt, resulting in an extra fetch too.
From bbolt docs:
// Seek positions the cursor at the passed seek key. If the key does not exist,
// the cursor is moved to the next key after seek. Returns the new pair.
