Mempool life cycle


The primary mempool object itself is initialized onto the node in init.cpp as part of AppInitMain() which takes NodeContext& node as an argument.

int check_ratio = std::min<int>(std::max<int>(args.GetIntArg("-checkmempool", chainparams.DefaultConsistencyChecks() ? 1 : 0), 0), 1000000);
node.mempool = std::make_unique<CTxMemPool>(node.fee_estimator.get(), check_ratio);

The check_ratio, used to determine sanity checks, defaults to 0 for all networks except regtest, unless the checkmempool program option has been specified.

Sanity checking here refers to checking the consistency of the entire mempool every 1 in n times a new transaction is added, so is potentially computationally expensive to have enabled.
See CTxMemPool::Check() for more information on what the check does.

Loading a previous mempool

If the node has been run before then it might have some blocks and a mempool to load. "Step 11: import blocks" of AppInitMain() in init.cpp calls ThreadImport() to load the mempool from disk where it is saved to file mempool.dat:

    chainman.m_load_block = std::thread(&TraceThread<std::function<void()>>, "loadblk", [=, &chainman, &args] {
        ThreadImport(chainman, vImportFiles, args);
This is run in its own thread so that (potentially) slow disk I/O has a minimal impact on startup times, and the remainder of startup execution can be continued.

ThreadImport runs a few jobs sequentially:

  1. Optionally perform a reindex

  2. Load the block files from disk

  3. Check that we are still on the best chain according to the blocks loaded from disk

  4. Load the mempool via chainman.ActiveChainstate().LoadMempool(args);

validation.cpp#LoadMempool() is an almost mirror of DumpMempool() described in more detail below in Mempool shutdown:

  1. Read the version and count of serialized transactions to follow

  2. Test each tx for expiry before submitting it to MemPoolAccept

  3. Read any remaining mapDeltas and unbroadcast_txids from the file and apply them

We test for expiry because it is current default policy not to keep transactions in the mempool longer than 336 hours, i.e. two weeks.
The default value comes from the constant DEFAULT_MEMPOOL_EXPIRE which can be overridden by the user with the -mempoolexpiry option.
Loading (and validating) a mempool of transactions this old is likely a waste of time and resources.

Runtime execution

While the node is running the mempool is persisted in memory. By default the mempool is limited to 300MB as specified by DEFAULT_MAX_MEMPOOL_SIZE. This can be overridden by the program option maxmempoolsize.

See mempool tx format for more information on what data counts towards this limit, or review the CTxMemPool data members which store current usage metrics e.g. CTxMemPool::cachedInnerUsage and the implementation of e.g. CTxMemPool::DynamicMemoryUsage().

Mempool shutdown

When the node is shut down its mempool is (by default) persisted to disk, called from init.cpp#Shutdown():

    if (node.mempool && node.mempool->IsLoaded() && node.args->GetArg("-persistmempool", DEFAULT_PERSIST_MEMPOOL)) {

A pointer to the mempool object is passed to DumpMempool(), which begins by locking the mempool mutex, pool.cs, before a snapshot of the mempool is created using local variables mapDeltas, vinfo and unbroadcast_txids.

mapDeltas is used by miners to apply (fee) prioritisation to certain transactions when creating new block templates.
vinfo stores information on each transaction as a vector of CTxMemPoolInfo objects.
bool DumpMempool(const CTxMemPool& pool, FopenFn mockable_fopen_function, bool skip_file_commit)
    int64_t start = GetTimeMicros();

    std::map<uint256, CAmount> mapDeltas;
    std::vector<TxMempoolInfo> vinfo;
    std::set<uint256> unbroadcast_txids;

    static Mutex dump_mutex;

        for (const auto &i : pool.mapDeltas) {
            mapDeltas[i.first] = i.second;
        vinfo = pool.infoAll();
        unbroadcast_txids = pool.GetUnbroadcastTxs();

Next a new (temporary) file is opened and some metadata related to mempool version and size is written to the front. Afterwards we loop through vinfo writing the transaction, the time it entered the mempool and the fee delta (prioritisation) to the file, before deleting its entry from our mapDeltas mirror.

Finally, any remaining info in mapDeltas is appended to the file. This might include prioritisation information on transactions not in our mempool.

    // ...
    try {
        FILE* filestr{mockable_fopen_function(GetDataDir() / "", "wb")};
        if (!filestr) {
            return false;

        CAutoFile file(filestr, SER_DISK, CLIENT_VERSION);

        uint64_t version = MEMPOOL_DUMP_VERSION;
        file << version;

        file << (uint64_t)vinfo.size();
        for (const auto& i : vinfo) {
            file << *(i.tx);
            file << int64_t{count_seconds(i.m_time)};
            file << int64_t{i.nFeeDelta};

        file << mapDeltas;

        LogPrintf("Writing %d unbroadcast transactions to disk.\n", unbroadcast_txids.size());
        file << unbroadcast_txids;
    // ...

We are able to write (and later read) mapDeltas and unbroadcast_txids to the file only using the << operator. This is due to the operator overload on the CAutoFile class found in streams.h:

 * map
template<typename Stream, typename K, typename T, typename Pred, typename A>
void Serialize(Stream& os, const std::map<K, T, Pred, A>& m)
    WriteCompactSize(os, m.size());
    for (const auto& entry : m)
        Serialize(os, entry);

class: CAutoFile
    // ...
    template<typename T>
    CAutoFile& operator<<(const T& obj)
        // Serialize to this stream
        if (!file)
            throw std::ios_base::failure("CAutoFile::operator<<: file handle is nullptr");
        ::Serialize(*this, obj);
        return (*this);
    // ...

Finally, if writing the elements to the temporary file was successful, we close the file and rename it to mempool.dat.