WIP Add LedgerStateCache #4642

ThomasBrady · 2025-02-06T07:03:34Z

Adds LedgerStateCache to store all Soroban entries:

Populates (LedgerStateCache::addEntry) the cache in PopulateLedgerCacheWork (executed in LedgerManagerImpl::loadLastKnownLedger).
Updates the cache (LedgerStateCache:::addEntries) each ledger from LedgerManagerImpl::transferLedgerEntriesToBucketList.
Reads the cache (LedgerStateCache::readEntry) in LedgerTxnRoot::Impl:::getNewestVersion.

Notes:

LedgerStateCache lives in an optional shared pointer and is created at startup, originally in the LedgerManagerImpl
LedgerTxnRoot takes an std::optional<LedgerStateCache> in its constructor, and PopulateLedgerCacheWork access it with LedgerManager::getLedgerStateCache.
addEntry and addEntries are private and only accessible from friend classes LedgerManagerImpl and PopulateLedgerCacheWork
addEntry and addEntries acquire a unique lock on the cache
Currently, the LedgerStateCache only supports contract entries. To access them in the index, I've added Bucket / Index::getContractEntryRange.
The LedgerStateCache is enabled via Config::IN_MEMORY_SOROBAN_STATE_FOR_TESTING (default true).

Description

Resolves #4556

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

…the cache in LedgerManagerImpl::loadLastKnownLedger via PopulateLedgerCacheWork. Updates the cache each ledger from LedgerManagerImpl::transferLedgerEntriesToBucketList. Reads the cache via LedgerTxnRoot::Impl:::getNewestVersion.

SirTyson

I think in general this looks good! I think we need explicit cache tests though. Specifically, we need population tests (which isn't covered in any unit tests currently I think). While we might have transitive coverage of actual cache usage, a few explicit unit tests would be useful as well.

I have two sorta high level comments. First, given that this is our base apply time cache and we're making protocol assumptions based on its efficiency, we need to be memory conscious and worry about things like copy semantics, probably preffering const shared_ptrs where we can.

2nd, I think we need to be more defensive in general about who has access with this cache. Many different threads and subsystems access state now after Marta'a background apply changes. However, each thread/subsystem has different expectations wrt that that access looks like. This cache in particular is an apply-time only cache, so I think it should be very limited to the apply time thread. I think to accomplish this we should have LedgerTxnRoot hold a unique_ptr to the cache instead of spreading around shared_ptrs. I think this is probably easiest if you construct a uniqe_ptr of the cache via your work class, then call some function like LedgerManager::setApplyTimeCache(unique_ptr&&). LedgerManager will then set the cache pointer in the ltx root. I think this is important as we want to make sure that only ltx has access to the cache. We might need to add an invalidation function too, such that the addbucketEntriesToCache function is a member of LedgerTxnRoot if that makes sense. This would also allow us to get rid of locks, since we'd be guaranteed everything is happening from the apply time thread. This might be different than what was discussed in the original design doc, but I think makes more sense after Marta's changes, where there is a more clear distinction between "apply time only state" and "state snapshots used by overlay and other subsystems"

SirTyson · 2025-02-06T22:24:10Z

src/bucket/InMemoryIndex.cpp

+                CLOG_DEBUG(Ledger, "Not a contract entry, type: {}", lk.type());
+            }
+        }
+        if (!lastContractEntry && firstContractEntry && lk.type() > TTL)


There is no entry larger than TTL, lastContractEntry.second is guarenteed to be EOF

SirTyson · 2025-02-06T22:24:46Z

src/bucket/InMemoryIndex.h

@@ -103,6 +105,12 @@ class InMemoryIndex
        return mOfferRange;
    }

+    std::optional<std::pair<std::streamoff, std::streamoff>>


I would change this to just return the starting offset, given that the ending offset is guaranteed to always be EOF

SirTyson · 2025-02-06T22:25:59Z

src/bucket/InMemoryIndex.cpp

+        // bound is EOF
+        else
+        {
+            CLOG_DEBUG(


These debug messages are super noisy, probably best to remove if tests are passing.

SirTyson · 2025-02-06T22:26:25Z

src/bucket/LiveBucket.cpp

+{
+    if (!getIndex().getContractEntryRange())
+    {
+        CLOG_DEBUG(Bucket, "LiveBucket::getContractEntryRange() = nullopt");


Nit: remove noisy message

SirTyson · 2025-02-06T22:27:39Z

src/bucket/LiveBucketIndex.cpp

+    if (mDiskIndex)
+    {
+        // Get the smallest and largest possible contract entry keys
+        LedgerKey upperBound(TTL /*9*/);


Upper bound not required since no entry is larger than TTL

SirTyson · 2025-02-06T23:17:58Z

src/ledger/LedgerStateCache.cpp

+        return;
+    }
+    mState.erase(entry);
+    mState.emplace(entry);


since this is population, we should just call emplace and assert it actually got inserted. There should be no duplicate inserts on construction.

SirTyson · 2025-02-06T23:21:43Z

src/ledger/LedgerManagerImpl.cpp

+        if (mLedgerStateCache)
+        {
+            auto populateLedgerCacheWork =
+                mApp.getWorkScheduler().executeWork<PopulateLedgerCacheWork>();


I think this would be better as a step in AssumeStateWork. That's where we do all our "in-memory matches current state" work, such as setting the BucketList, indexing buckets, etc.

SirTyson · 2025-02-06T23:23:46Z

src/ledger/LedgerTxn.cpp

@@ -534,6 +535,8 @@ LedgerTxn::Impl::commitChild(EntryIterator iter,
                             RestoredKeys const& restoredKeys,
                             LedgerTxnConsistency cons) noexcept
 {
+    // We will want to acquire the write lock to the cache in this function.


Remove comment

SirTyson · 2025-02-06T23:27:07Z

src/catchup/PopulateLedgerCacheWork.cpp

+            mBucketsToProcess.push_back(bucket);
+            CLOG_DEBUG(Ledger, "Adding bucket {} to mBucketsToProcess[{}]",
+                       xdr::xdr_to_string(bucket->getHash()), counter);
+            counter++;


Nit: ++counter

SirTyson · 2025-02-06T23:27:51Z

src/catchup/PopulateLedgerCacheWork.cpp

+        {
+            mBucketsToProcess.push_back(bucket);
+            CLOG_DEBUG(Ledger, "Adding bucket {} to mBucketsToProcess[{}]",
+                       xdr::xdr_to_string(bucket->getHash()), counter);


Nit: use binToHex() or hexAbbrev() instead of xdr_to_string

ThomasBrady requested a review from SirTyson February 6, 2025 17:51

Add non-soroban testonly ledgerstateconfig mode

8df0219

ThomasBrady force-pushed the global-in-memory-cache branch from c1f11fd to 8df0219 Compare February 6, 2025 22:29

SirTyson requested changes Feb 6, 2025

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

WIP Add LedgerStateCache #4642

WIP Add LedgerStateCache #4642

ThomasBrady commented Feb 6, 2025 •

edited

Loading

SirTyson left a comment

SirTyson Feb 6, 2025

SirTyson Feb 6, 2025

SirTyson Feb 6, 2025

SirTyson Feb 6, 2025

SirTyson Feb 6, 2025

SirTyson Feb 6, 2025

SirTyson Feb 6, 2025

SirTyson Feb 6, 2025

SirTyson Feb 6, 2025

SirTyson Feb 6, 2025

WIP Add LedgerStateCache #4642

Are you sure you want to change the base?

WIP Add LedgerStateCache #4642

Conversation

ThomasBrady commented Feb 6, 2025 • edited Loading

Description

Checklist

SirTyson left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

ThomasBrady commented Feb 6, 2025 •

edited

Loading