cache latest checkpoint in-memory to avoid deserializing and reserializing it too often

[KtorZ]

Issue Number

Overview

This little change has some drastic performance improvements on
various tasks of the wallet:

Wallet Overview
   number of addresses: 29209
   number of transactions: 48819
   number of utxos: 2778 UTxOs

Bench                   | No Cache | With Cache
---                     | --       | ---
restoreTime             | 63min    | 22min
readWalletTime          | 715ms    | 53.70 ms
listAddressesTime       | 1.863s   | 2.169 ms
listTransactionsTime    | 75.72s   | 12.68 s
importOneAddressTime    | 2.526s   | 260.4 ms
importManyAddressesTime | 3.104s   | 1.690 s
estimateFeesTime        | 2.103s   | 63.11 ms

This cache approach however supposes that there's only a single wallet
per database, which is the case and has been the case for a long time
now. I'll remove the WalletId from the database interface in a next
commit to clean things up a bit.

Comments

[Anviking]

The cache doesn't live in atomically, in contrast to the rest of the DB layer. Couldn't this lead to race-conditions?

[KtorZ]

There can only be a single execution thread in atomically (constructed with an MVar lock), hence the use of a mere IORef here.

[Anviking]

I was imagining something more like:

----------------------------------------------------------------------> time
1: putCheckpoint cp1 
2:           selectLatest >>= f >>= putCheckpoint
                                                     selectLatest >>= g >>= putCheckpoint

1. Updates cache, but operation is cancelled before the DB is updated. Cache and DB diverges.
2. DB cp will be overwritten by the cached checkpoint the next time it is modified

But I guess the worst thing that can happen is that a somehow-failing putCheckpoint will end up succeeding 🤔

[KtorZ]

Updates cache, but operation is cancelled before the DB is updated. Cache and DB diverges.

Cache is updated after the put operation has succeeded though. If it doesn't succeed, the cache isn't updated.

[Anviking]

Ok, Matthias pointed out that "We create an illusion of transaction by having a top-level MVar lock.", so there shouldn't be Sqlite transactions being run and then somehow cancelled.

[KtorZ]

I also added some notes in the code about this in place where the cache is created.

[Anviking]

😅 Perhaps worth adding to .gitignore
?

[KtorZ]

Whoops. Sounds like I a wrong git add lib indeed :)

[Anviking]

Nice speedup. LGTM.

[Anviking]

Is there a chance of buildup of thunks here? Though I suppose we'd notice that when testing this

[KtorZ]

Thought of it and:

• Map is strict, so values are evaluated before they are put in the Map
• We never just write in the cache without reading and evaluating data from it first.

So no, I don't think there's any chance that large thunks will be built here. I could switch to modifyIORef' for the tranquility of mind, we don't need the lazyness anyway since we have in principle already evaluated the cache to WHNF (and that's actually part of why we want to cache it!)

[KtorZ]

bors try

[iohk-bors]

try

Build failed:

• ci/hydra-build:required

[KtorZ]

bors r+

[iohk-bors]

Timed out.
#2216

[KtorZ]

bors retry

[iohk-bors]

Build failed:

• ci/hydra-build:required

Log-limit exceeded, perhaps due to running out of faucet funds, which should be fixed on recent master.
#2216

[KtorZ]

bors r+

[iohk-bors]

Build failed:

• ci/hydra-build:required

#2218 ADP-619

  src/Test/Integration/Scenario/API/Byron/Addresses.hs:410:46:
  1) API Specifications, BYRON_ADDRESSES, ADDRESS_IMPORT_05 - I can import 15000 of addresses
       uncaught exception: IOException of type UserError
       user error (Waited longer than 90s (more than 2 epochs) for an action to resolve. Action: "Addresses are imported". Error condition: Just (HUnitFailure (Just (SrcLoc {srcLocPackage = "cardano-wallet-core-integration-2020.9.22-6DPlnrdhceJLXA8CsrMoUY", srcLocModule = "Test.Integration.Scenario.API.Byron.Addresses", srcLocFile = "src/Test/Integration/Scenario/API/Byron/Addresses.hs", srcLocStartLine = 430, srcLocStartCol = 13, srcLocEndLine = 430, srcLocEndCol = 35})) (ExpectedButGot Nothing "15000" "0")))

  To rerun use: --match "/API Specifications/BYRON_ADDRESSES/ADDRESS_IMPORT_05 - I can import 15000 of addresses/"

[KtorZ]

bors merge

[iohk-bors]

Build succeeded:

• ci/hydra-build:required
• buildkite/cardano-wallet