Introduction

≈60 min · Advanced · builds on Build & run your own xrpld node

Now that your node runs, let's see what actually holds it together. At the center of rippled sits one object (the Application class) that owns and wires up every subsystem, from the ledger to the network to storage. In this module you'll trace how the node starts up, runs and shuts down, meet the major subsystems and the job queue that schedules their work, and learn how any component reaches another. It's the map you'll keep coming back to as the codebase gets deeper.

The Application Class Architecture

In brief: the single object that owns every subsystem and hands them out on request (a service locator).

Design Philosophy

The Application class follows several key design principles that make Rippled maintainable and extensible:

Single Point of Coordination: Instead of components directly creating and managing their dependencies, everything flows through the Application. This centralization makes it easy to understand system initialization and component relationships.

Dependencies through the Application: rippled predominantly uses the service locator style: subsystems receive Application& and ask it for what they need (app.getLedgerMaster()), rather than receiving each dependency in their constructor. A few components do take explicit constructor dependencies, but Application& is the norm; it trades some testability for one obvious place to find everything.

Interface-Based Design: The Application class implements the Application interface, allowing for different implementations (production, test, mock) without changing dependent code. The subsystem accessors themselves live on a separate ServiceRegistry interface that Application inherits, so components that only need service access (not lifecycle control) can hold a ServiceRegistry& instead.

Lifetime Management: The Application controls the creation, initialization, and destruction of all major subsystems, ensuring proper startup/shutdown sequences.

Application Interface

The Application interface is defined in src/xrpld/app/main/Application.h. It inherits the subsystem accessors from ServiceRegistry (include/xrpl/core/ServiceRegistry.h) and adds lifecycle control plus a few utilities:

class Application : public ServiceRegistry, public beast::PropertyStream::Source
{
public:
    using MutexType = std::recursive_mutex;
    virtual MutexType&
    getMasterMutex() = 0;

public:
    Application();

    virtual bool
    setup(boost::program_options::variables_map const& options) = 0;

    virtual void
    start(bool withTimers) = 0;
    virtual void
    run() = 0;
    virtual void
    signalStop(std::string msg) = 0;

    // ...

    virtual Config&
    config() = 0;

    // ...

    /** Ensure that a newly-started validator does not sign proposals older
     * than the last ledger it persisted. */
    virtual LedgerIndex
    getMaxDisallowedLedger() = 0;

    // ...
};

The interface every subsystem is reached through is ServiceRegistry (trimmed to the accessors you'll use most):

class ServiceRegistry
{
public:
    // ...
    virtual JobQueue&
    getJobQueue() = 0;

    // ...

    virtual RCLValidations&
    getValidations() = 0;

    // ...

    virtual ManifestCache&
    getValidatorManifests() = 0;

    virtual ManifestCache&
    getPublisherManifests() = 0;

    // Network services
    virtual Overlay&
    getOverlay() = 0;

    // ...

    // Storage services
    virtual NodeStore::Database&
    getNodeStore() = 0;

    // ...

    virtual RelationalDatabase&
    getRelationalDatabase() = 0;

    // Ledger services
    virtual InboundLedgers&
    getInboundLedgers() = 0;

    // ...

    virtual LedgerMaster&
    getLedgerMaster() = 0;

    // ...

    virtual OpenLedger&
    getOpenLedger() = 0;

    // ...

    // Transaction and operation services
    virtual NetworkOPs&
    getOPs() = 0;

    // ...

    virtual TxQ&
    getTxQ() = 0;

    // ...

    // Server services
    virtual ServerHandler&
    getServerHandler() = 0;

    // ...

    // Configuration and state
    [[nodiscard]] virtual bool
    isStopping() const = 0;

    // ...

    virtual Logs&
    getLogs() = 0;

    // ...
};

Note there is no RPCHandler class: RPC commands are dispatched through free functions in src/xrpld/rpc/RPCHandler.h (RPC::doCommand), and the RPC/WebSocket servers are managed by ServerHandler, reached via getServerHandler().

ApplicationImp Implementation

The concrete implementation ApplicationImp is in src/xrpld/app/main/Application.cpp. This class:

Implements all interface methods
Owns all major subsystem objects
Manages initialization order
Coordinates shutdown
Provides cross-cutting services

Key Member Variables (trimmed; note they are public, under a NOLINT block, and several are held by value or std::optional rather than unique_ptr):

class ApplicationImp : public Application, public BasicApp
{
public:
    // NOLINTBEGIN(readability-identifier-naming)
    std::unique_ptr<Config> config_;
    std::unique_ptr<Logs> logs_;
    std::unique_ptr<TimeKeeper> timeKeeper_;

    // ...

    beast::Journal journal_;
    std::unique_ptr<perf::PerfLog> perfLog_;

    // ...

    std::unique_ptr<JobQueue> jobQueue_;

    // ...

    std::unique_ptr<SHAMapStore> shaMapStore_;

    // ...

    std::optional<OpenLedger> openLedger_;

    // ...

    std::unique_ptr<NodeStore::Database> nodeStore_;

    // ...

    std::unique_ptr<LedgerMaster> ledgerMaster_;
    std::unique_ptr<LedgerCleaner> ledgerCleaner_;
    std::unique_ptr<InboundLedgers> inboundLedgers_;
    std::unique_ptr<InboundTransactions> inboundTransactions_;

    // ...

    std::unique_ptr<NetworkOPs> networkOPs_;

    // ...

    std::unique_ptr<ServerHandler> serverHandler_;
    std::unique_ptr<AmendmentTable> amendmentTable_;

    // ...

    RCLValidations validations_;
    std::unique_ptr<LoadManager> loadManager_;
    std::unique_ptr<TxQ> txQ_;

    // ...

    std::optional<SQLiteDatabase> relationalDatabase_;
    std::unique_ptr<DatabaseCon> walletDB_;
    std::unique_ptr<Overlay> overlay_;

    // ...

    std::atomic_flag isTimeToStop;

    // ...

    std::unique_ptr<GRPCServer> grpcServer_;
    // NOLINTEND(readability-identifier-naming)
};

Key idea. Almost nothing in rippled constructs its own dependencies; it asks the Application for them. Find the Application and you can reach the whole node.

Initialization and Lifecycle

In brief: how the node starts up, runs, and shuts down, in order.

Startup Sequence

Understanding the startup sequence is crucial for debugging initialization issues and understanding component dependencies. The real call chain in src/xrpld/app/main/Main.cpp is: load Config → makeApplication(...) → app->setup(vm) → app->start(true) → app->run().

Phase 1: Configuration Loading

// In Main.cpp
auto config = std::make_unique<Config>();

auto configFile = vm.contains("conf") ? vm["conf"].as<std::string>() : std::string();

// config file, quiet flag.
config->setup(
    configFile, vm.contains("quiet"), vm.contains("silent"), vm.contains("standalone"));

What Happens:

Parse xrpld.cfg configuration file
Load validator list configuration
Set up logging configuration
Validate configuration parameters
Apply defaults for unspecified options

Configuration Sections:

[server] - Server ports and interfaces
[node_db] - NodeStore database configuration
[node_size] - Performance tuning parameters
[validation_seed] - Validator key configuration
[ips_fixed] - Fixed peer connections
[features] - Amendments to treat as enabled locally (this is not voting)
[amendments] / [veto_amendments] - Amendment voting (up-vote / down-vote)

Phase 2: Application Construction

// Create the application instance
auto app =
    makeApplication(std::move(config), std::move(logs), std::make_unique<TimeKeeper>());

Constructor Sequence (ApplicationImp::ApplicationImp()): the subsystems are constructed in the member-initializer list, in member declaration order — not assigned in the constructor body. Trimmed:

ApplicationImp(
    std::unique_ptr<Config> config,
    std::unique_ptr<Logs> logs,
    std::unique_ptr<TimeKeeper> timeKeeper)
    : BasicApp(numberOfThreads(*config))
    , config_(std::move(config))
    , logs_(std::move(logs))
    , timeKeeper_(std::move(timeKeeper))
    // ...
    , journal_(logs_->journal("Application"))
    // PerfLog must be started before any other threads are launched.
    , perfLog_(
          perf::makePerfLog(
              perf::setupPerfLog(config_->section("perf"), config_->configDir),
              *this,
              logs_->journal("PerfLog"),
              [this] { signalStop("PerfLog"); }))
    , txMaster_(*this)
    // ...
    , jobQueue_(std::make_unique<JobQueue>(/* thread count, collector, journals... */))
    , nodeStoreScheduler_(*jobQueue_)
    , shaMapStore_(makeSHAMapStore(*this, nodeStoreScheduler_, logs_->journal("SHAMapStore")))
    // ...
    , nodeStore_(shaMapStore_->makeNodeStore(
          config_->prefetchWorkers > 0 ? config_->prefetchWorkers : 4))
    // ...
    , ledgerMaster_(
          std::make_unique<LedgerMaster>(
              *this,
              stopwatch(),
              collectorManager_->collector(),
              logs_->journal("LedgerMaster")))
    // ...
    , networkOPs_(makeNetworkOPs(/* ... */))
    // ...
    , validations_(ValidationParms(), stopwatch(), *this, logs_->journal("Validations"))
    , loadManager_(makeLoadManager(*this, logs_->journal("LoadManager")))
    , txQ_(std::make_unique<TxQ>(setupTxQ(*config_), logs_->journal("TxQ")))
    // ...
{
    // ...
    //  Do not start threads, open sockets, or do any sort of "real work"
    //  inside the constructor. Put it in start instead. Or if you must,
    //  put it in setup (but everything in setup should be moved to start
    //  anyway.
    // ...
}

Notes on the real order:

The first subsystem is PerfLog, with the comment "PerfLog must be started before any other threads are launched."
The NodeStore is created through shaMapStore_->makeNodeStore(...), not directly.
The Overlay is not created here — it stays a null unique_ptr until setup().
The relational database is also opened later, in setup(), via initRelationalDatabase().
Order matters! Components may depend on earlier ones.

Phase 3: Setup

if (!app->setup(vm))
    return -1;

What Happens (ApplicationImp::setup() returns bool; trimmed):

bool
ApplicationImp::setup(boost::program_options::variables_map const& cmdline)
{
    // ...

    if (!initRelationalDatabase() || !initNodeStore())
        return false;

    // ...

    // Configure the amendments the server supports
    {
        // ...
        Section const& downVoted = config_->section(SECTION_VETO_AMENDMENTS);

        Section const& upVoted = config_->section(SECTION_AMENDMENTS);

        amendmentTable_ = makeAmendmentTable(
            *this,
            config().amendmentMajorityTime,
            supported,
            upVoted,
            downVoted,
            logs_->journal("Amendments"));
    }

    // ...

    overlay_ = makeOverlay(
        *this,
        setupOverlay(*config_, journal_),
        *serverHandler_,
        *resourceManager_,
        *resolver_,
        getIoContext(),
        *config_,
        collectorManager_->collector());
    add(*overlay_);  // add to PropertyStream

    // start first consensus round
    if (!networkOPs_->beginConsensus(ledgerMaster_->getClosedLedger()->header().hash, {}))
    {
        JLOG(journal_.fatal()) << "Unable to start consensus";
        return false;
    }

    // ...
}

Setup also loads the last ledger state (or creates a genesis ledger in standalone mode), loads peer reservations, manifests and validator lists, and configures the ServerHandler ports. Note that amendment voting comes from the [amendments] and [veto_amendments] config sections here.

Phase 4: Start

// Start the server
app->start(true /*start timers*/);

What Happens (ApplicationImp::start(bool withTimers)):

void
ApplicationImp::start(bool withTimers)
{
    JLOG(journal_.info()) << "Application starting. Version is " << BuildInfo::getVersionString();

    if (withTimers)
    {
        setSweepTimer();
        setEntropyTimer();
    }

    io_latency_sampler_.start();
    resolver_->start();
    loadManager_->start();
    shaMapStore_->start();
    if (overlay_)
        overlay_->start();

    if (grpcServer_->start())
        fixConfigPorts(*config_, {{SECTION_PORT_GRPC, grpcServer_->getEndpoint()}});

    ledgerCleaner_->start();
    perfLog_->start();
}

This is where threads, sockets, and timers actually come alive — exactly the "real work" the constructor comment forbids.

Phase 5: Run

// Block until we get a stop RPC.
app->run();

Main Event Loop (ApplicationImp::run(), trimmed):

void
ApplicationImp::run()
{
    if (!config_->standalone())
    {
        // ...
        getLoadManager().activateStallDetector();
    }

    isTimeToStop.wait(false, std::memory_order_relaxed);

    JLOG(journal_.debug()) << "Application stopping";

    // ... (the io latency sampler, resolver, and the sweep/entropy
    //      timers are cancelled; manifests are saved to wallet.db) ...

    // The order of these stop calls is delicate.
    // Re-ordering them risks undefined behavior.
    loadManager_->stop();
    shaMapStore_->stop();
    jobQueue_->stop();
    if (overlay_)
        overlay_->stop();
    grpcServer_->stop();
    networkOPs_->stop();
    serverHandler_->stop();
    ledgerReplayer_->stop();
    inboundTransactions_->stop();
    inboundLedgers_->stop();
    ledgerCleaner_->stop();
    nodeStore_->stop();
    perfLog_->stop();

    JLOG(journal_.info()) << "Done.";
}

What Runs:

Job queue processes queued work
Overlay network handles peer connections
Consensus engine processes rounds
NetworkOPs coordinates operations
RPC handlers process client requests

All work happens in background threads managed by various subsystems. The main thread simply blocks in run() on the isTimeToStop atomic flag until a shutdown is signalled — and then run() itself performs the teardown shown above.

Phase 6: Shutdown

// e.g. in the "stop" RPC handler
// (src/xrpld/rpc/handlers/admin/server_control/Stop.cpp)
context.app.signalStop("RPC");

signalStop() does not perform the shutdown. Its entire job is to set the stop flag and wake the main thread (ApplicationImp::signalStop()):

void
ApplicationImp::signalStop(std::string msg)
{
    if (!isTimeToStop.test_and_set(std::memory_order_acquire))
    {
        if (msg.empty())
        {
            JLOG(journal_.warn()) << "Server stopping";
        }
        else
            JLOG(journal_.warn()) << "Server stopping: " << msg;

        isTimeToStop.notify_all();
    }
}

It is called from the stop admin RPC, from the POSIX signal handler (signalStop("Signal: " + to_string(signum))), and even from subsystems themselves (for example when the transaction database runs out of space). The actual teardown then runs at the end of run().

Shutdown Order: The stop calls at the end of run() are not simply the reverse of construction order — jobQueue_ is among the first members constructed yet stops third, while overlay_ is created last (in setup()) yet stops fourth. The code's own comment says it best: "The order of these stop calls is delicate. Re-ordering them risks undefined behavior."

Complete Lifecycle Diagram

The application lifecycle: program start, load configuration (xrpld.cfg), create the Application instance, construct the subsystems, setup phase, run phase (the main loop and steady state), shutdown signal, graceful shutdown, program exit

Subsystem Coordination

In brief: how components find and call each other through the Application.

The Service Locator Pattern

The Application acts as a service locator, allowing any component to access any other component through the app reference:

class SomeComponent
{
public:
    SomeComponent(Application& app)
        : app_(app)
    {
        // Components store app reference
    }

    void doWork()
    {
        // Access other components through app
        auto& ledgerMaster = app_.getLedgerMaster();
        auto& overlay = app_.getOverlay();
        auto& jobs = app_.getJobQueue();

        // Use the components...
    }

private:
    Application& app_;
};

Major Subsystems

LedgerMaster

Purpose: Manages the chain of validated ledgers and coordinates ledger progression.

Key Responsibilities:

Track current validated ledger
Build candidate ledgers for consensus
Synchronize ledger history
Maintain ledger cache
Coordinate with consensus engine

Access: app.getLedgerMaster()

Important Methods (from src/xrpld/app/ledger/LedgerMaster.h):

// The finalized ledger is the last closed/accepted ledger
std::shared_ptr<Ledger const>
getClosedLedger()
{
    return closedLedger_.get();
}

// The validated ledger is the last fully validated ledger.
std::shared_ptr<Ledger const>
getValidatedLedger();

void
tryAdvance();

tryAdvance() is the entry point for advancing the validated ledger (it schedules a JtAdvance job). Fetching missing ledgers is not LedgerMaster's job: that belongs to InboundLedgers, reached via app.getInboundLedgers().

NetworkOPs

Purpose: Coordinates network operations and transaction processing.

Key Responsibilities:

Process submitted transactions
Manage transaction queue
Coordinate consensus participation
Track network state
Publish ledger close events

Access: app.getOPs()

Important Methods (from include/xrpl/server/NetworkOPs.h):

// must complete immediately
virtual void
submitTransaction(std::shared_ptr<STTx const> const&) = 0;

/**
 * Process transactions as they arrive from the network or which are
 * submitted by clients. Process local transactions synchronously
 *
 * @param transaction Transaction object
 * @param bUnlimited Whether a privileged client connection submitted it.
 * @param bLocal Client submission.
 * @param failType fail_hard setting from transaction submission.
 */
virtual void
processTransaction(
    std::shared_ptr<Transaction>& transaction,
    bool bUnlimited,
    bool bLocal,
    FailHard failType) = 0;

[[nodiscard]] virtual OperatingMode
getOperatingMode() const = 0;

Overlay

Purpose: Manages peer-to-peer networking layer.

Key Responsibilities:

Peer discovery and connection
Message routing
Network topology maintenance
Bandwidth management

Access: app.getOverlay()

Important Methods (from src/xrpld/overlay/Overlay.h):

/** Establish a peer connection to the specified endpoint.
    The call returns immediately, the connection attempt is
    performed asynchronously.
*/
virtual void
connect(beast::IP::Endpoint const& address) = 0;

/** Returns the number of active peers.
    Active peers are only those peers that have completed the
    handshake and are using the peer protocol.
*/
[[nodiscard]] virtual std::size_t
size() const = 0;

/** Broadcast a proposal. */
virtual void
broadcast(protocol::TMProposeSet& m) = 0;

/** Broadcast a validation. */
virtual void
broadcast(protocol::TMValidation& m) = 0;

Note there is no generic broadcast(Message): only proposals and validations are broadcast to every peer; other traffic goes through relay/send paths.

TxQ (Transaction Queue)

Purpose: Manages transaction queuing when network is busy.

Key Responsibilities:

Queue transactions during high load
Fee-based prioritization
Account-based queuing limits
Transaction expiration

Access: app.getTxQ()

Important Methods (from src/xrpld/app/misc/TxQ.h):

/**
    Add a new transaction to the open ledger, hold it in the queue,
    or reject it.

    @return A pair with the `TER` and a `bool` indicating
            whether or not the transaction was applied to
            the open ledger. If the transaction is queued,
            will return `{ terQUEUED, false }`.
*/
ApplyResult
apply(
    Application& app,
    OpenView& view,
    std::shared_ptr<STTx const> const& tx,
    ApplyFlags flags,
    beast::Journal j);

NodeStore

Purpose: Persistent storage for ledger data.

Key Responsibilities:

Store ledger state nodes
Provide efficient retrieval
Cache frequently accessed data
Support different backend databases (RocksDB, NuDB)

Access: app.getNodeStore()

Important Methods (from include/xrpl/nodestore/Database.h):

/** Store the object.

    The caller's Blob parameter is overwritten.

    @param type The type of object.
    @param data The payload of the object. The caller's
                variable is overwritten.
    @param hash The 256-bit hash of the payload data.
    @param ledgerSeq The sequence of the ledger the object belongs to.

    @return `true` if the object was stored?
*/
virtual void
store(NodeObjectType type, Blob&& data, uint256 const& hash, std::uint32_t ledgerSeq) = 0;

/** Fetch a node object.
    If the object is known to be not in the database, isn't found in the
    database during the fetch, or failed to load correctly during the fetch,
    `nullptr` is returned.

    @note This can be called concurrently.
    @param hash The key of the object to retrieve.
    @param ledgerSeq The sequence of the ledger where the object is stored.
    @param fetchType the type of fetch, synchronous or asynchronous.
    @return The object, or nullptr if it couldn't be retrieved.
*/
std::shared_ptr<NodeObject>
fetchNodeObject(
    uint256 const& hash,
    std::uint32_t ledgerSeq = 0,
    FetchType fetchType = FetchType::Synchronous,
    bool duplicate = false);

RelationalDatabase

Purpose: SQL database for indexed data and historical queries.

Key Responsibilities:

Store transaction metadata
Maintain account transaction history
Support RPC queries (account_tx, tx)
Ledger header storage

Access: app.getRelationalDatabase()

Database Types:

SQLite (embedded) — the only relational backend. ApplicationImp holds a std::optional<SQLiteDatabase> directly; the PostgreSQL backend was removed together with reporting mode.

Validations

Purpose: Manages validator signatures on ledger closes.

Key Responsibilities:

Collect validations from validators
Track which validations are current, trusted, and full
Count trusted validations for a ledger (the raw input to validation decisions)
Publish validation stream

Access: app.getValidations() — returns RCLValidations&, an alias for Validations<RCLValidationsAdaptor> (the generic engine in src/xrpld/consensus/Validations.h with an RCL-specific adaptor).

Important Methods (from src/xrpld/consensus/Validations.h, trimmed to declarations):

/** Add a new validation

    Attempt to add a new validation.

    @param nodeID The identity of the node issuing this validation
    @param val The validation to store
    @return The outcome
*/
ValStatus
add(NodeID const& nodeID, Validation const& val);

/** Get the currently trusted full validations

    @return Vector of validations from currently trusted validators
*/
std::vector<WrappedValidationType>
currentTrusted();

/** Count the number of trusted full validations for the given ledger

    @param ledgerID The identifier of ledger of interest
    @return The number of trusted validations
*/
std::size_t
numTrustedForLedger(ID const& ledgerID);

/**  Get trusted full validations for a specific ledger

     @param ledgerID The identifier of ledger of interest
     @param seq The sequence number of ledger of interest
     @return Trusted validations associated with ledger
*/
std::vector<WrappedValidationType>
getTrustedForLedger(ID const& ledgerID, Seq const& seq);

Incoming validations from the network enter through the free function handleNewValidation (src/xrpld/app/consensus/RCLValidations.h):

void
handleNewValidation(
    Application& app,
    std::shared_ptr<STValidation> const& val,
    std::string const& source,
    BypassAccept const bypassAccept = BypassAccept::No,
    std::optional<beast::Journal> j = std::nullopt);

Two related responsibilities live elsewhere: validator key rotations (manifests) are tracked by ManifestCache, reached via app.getValidatorManifests() and app.getPublisherManifests(), and the quorum policy (how many trusted validators are needed) lives in ValidatorList, reached via app.getValidators().

Job Queue System

In brief: how background work is queued, prioritized, and run across worker threads.

Purpose and Design

The job queue is Rippled's work scheduling system. Instead of each subsystem creating its own threads, work is submitted as jobs to a centralized queue processed by a thread pool. This provides:

Centralized thread management: Easier to control thread count and CPU usage
Priority-based scheduling: Critical jobs processed before low-priority ones
Visibility: Easy to monitor what work is queued
Deadlock prevention: Structured concurrency patterns

Job Types

Jobs are categorized by type, which determines priority. From include/xrpl/core/Job.h (trimmed) — read the comment carefully: earlier entries have LOWER priority, so the highest-priority job types sit at the bottom of the enum:

// NOLINTNEXTLINE(cppcoreguidelines-use-enum-class)
enum JobType {
    // Special type indicating an invalid job - will go away soon.
    JtInvalid = -1,

    // Job types - the position in this enum indicates the job priority with
    // earlier jobs having lower priority than later jobs. If you wish to
    // insert a job at a specific priority, simply add it at the right location.

    JtPack,             // Make a fetch pack for a peer
    JtPuboldledger,     // An old ledger has been accepted
    JtClient,           // A placeholder for the priority of all jtCLIENT jobs
    JtClientSubscribe,  // A websocket subscription by a client
    // ... (more client job types) ...
    JtRpc,              // A websocket command from the client
    JtSweep,            // Sweep for stale structures
    JtValidationUt,     // A validation from an untrusted source
    JtManifest,         // A validator's manifest
    JtUpdatePf,         // Update pathfinding requests
    JtTransactionL,     // A local transaction
    // ...
    JtProposalUt,       // A proposal from an untrusted source
    // ...
    JtTransaction,      // A transaction received from the network
    // ...
    JtBatch,            // Apply batched transactions
    JtLedgerData,       // Received data for a ledger we're acquiring
    JtAdvance,          // Advance validated/acquired ledgers
    JtPubledger,        // Publish a fully-accepted ledger
    JtTxnData,          // Fetch a proposed set
    JtWal,              // Write-ahead logging
    JtValidationT,      // A validation from a trusted source
    JtWrite,            // Write out hashed objects
    JtAccept,           // Accept a consensus ledger
    JtProposalT,        // A proposal from a trusted source
    JtNetopCluster,     // NetworkOPs cluster peer report
    JtNetopTimer,       // NetworkOPs net timer processing
    JtAdmin,            // An administrative operation

    // Special job types which are not dispatched by the job pool
    JtPeer,
    JtDisk,
    // ...
};

So the consensus-critical types — JtValidationT (trusted validations), JtAccept (accept a consensus ledger), and JtProposalT (trusted proposals) — are near the bottom, i.e. among the highest priorities, with JtAdmin the highest-priority dispatched type. JtPack and JtPuboldledger at the top are the two lowest. JtPeer and JtDisk are never queued at all; they exist only for load measurement.

Submitting Jobs

Components submit work to the job queue via addJob. The handler is a lambda taking no arguments — the addJob template (include/xrpl/core/JobQueue.h) enforces this at compile time:

template <
    typename JobHandler,
    typename = std::enable_if_t<std::is_same_v<decltype(std::declval<JobHandler&&>()()), void>>>
bool
addJob(JobType type, std::string const& name, JobHandler&& jobHandler)

A real call, from NetworkOPsImp::submitTransaction (src/xrpld/app/misc/NetworkOPs.cpp):

jobQueue_.addJob(JtTransaction, "SubmitTxn", [this, tx]() {
    auto t = tx;
    processTransaction(t, false, false, FailHard::No);
});

(Outside NetworkOPs, get the queue with app.getJobQueue() first.)

Job Priority and Scheduling

Priority Levels (from the real enum order — higher in this list = processed first):

Highest: JtAdmin (administrative operations), then the NetworkOPs timer/cluster jobs
Consensus-critical: JtProposalT, JtAccept, JtValidationT (trusted proposals, ledger accept, trusted validations)
Ledger progression: JtPubledger, JtAdvance, JtLedgerData, JtTransaction
Lower: untrusted validations/proposals, pathfinding updates, sweeps
Lowest dispatched: client and RPC jobs (JtClient*, JtRpc), JtPuboldledger, JtPack

Scheduling Algorithm:

Worker threads pick the highest-priority waiting job
Each job type also has a per-type queue limit (include/xrpl/core/JobTypes.h) — this, not priority alone, is what protects the node under load
Each type carries target average/peak latencies used for load measurement

A slice of the real table (JobTypes.h):

//                                                           avg     peak
//  JobType               name                    limit    latency  latency
add(JtPack,              "makeFetchPack",               1,     0ms,     0ms);
add(JtPuboldledger,      "publishAcqLedger",            2, 10000ms, 15000ms);
add(JtValidationUt,     "untrustedValidation",  maxLimit,  2000ms,  5000ms);
add(JtLedgerData,       "ledgerData",                  3,     0ms,     0ms);
add(JtUpdatePf,         "updatePaths",                 1,     0ms,     0ms);
add(JtTransaction,       "transaction",          maxLimit,   250ms,  1000ms);
add(JtAdvance,           "advanceLedger",        maxLimit,     0ms,     0ms);
add(JtPubledger,         "publishNewLedger",     maxLimit,  3000ms,  4500ms);
add(JtValidationT,      "trustedValidation",    maxLimit,   500ms,  1500ms);
add(JtAccept,            "acceptLedger",         maxLimit,     0ms,     0ms);
add(JtProposalT,        "trustedProposal",      maxLimit,   100ms,   500ms);
add(JtAdmin,             "administration",       maxLimit,     0ms,     0ms);

The name column is what you'll see in the logs (e.g. transaction, advanceLedger, publishNewLedger), not the enum constant.

Job Queue Configuration

In xrpld.cfg:

[node_size]
# Influences the worker thread count (and many cache sizes).
# The default is tiny.
medium

[workers]
# Optional: explicitly set the job queue thread count.

The real thread-count logic is the lambda that constructs the JobQueue in ApplicationImp's member-initializer list (src/xrpld/app/main/Application.cpp):

[](std::unique_ptr<Config> const& config) {
    if (config->standalone() && !config->forceMultiThread)
        return 1;

    if (config->workers)
        return config->workers;

    auto count = static_cast<int>(std::thread::hardware_concurrency());

    // Be more aggressive about the number of threads to use
    // for the job queue if the server is configured as
    // "large" or "huge" if there are enough cores.
    if (config->nodeSize >= 4 && count >= 16)
    {
        count = 6 + std::min(count, 8);
    }
    else if (config->nodeSize >= 3 && count >= 8)
    {
        count = 4 + std::min(count, 6);
    }
    else
    {
        count = 2 + std::min(count, 4);
    }

    return count;
}(config_)

So: 1 thread in standalone mode, the [workers] value if set, otherwise a core-count-derived number that only grows for large/huge node sizes on machines with enough cores.

Configuration Management

In brief: how the node reads its config and exposes it to every subsystem.

Configuration File Structure

The xrpld.cfg file controls all aspects of server behavior. The Application loads and provides access to this configuration.

Example Configuration

[server]
port_rpc_admin_local
port_peer
port_ws_admin_local

[port_rpc_admin_local]
port = 5005
ip = 127.0.0.1
admin = 127.0.0.1
protocol = http

[port_peer]
port = 51235
ip = 0.0.0.0
protocol = peer

[port_ws_admin_local]
port = 6006
ip = 127.0.0.1
admin = 127.0.0.1
protocol = ws

[node_size]
medium

[node_db]
type=RocksDB
path=/var/lib/rippled/db/rocksdb
open_files=512
cache_mb=256
filter_bits=12
compression=1

[database_path]
/var/lib/rippled/db

[debug_logfile]
/var/log/rippled/debug.log

[sntp_servers]
time.windows.com
time.apple.com
time.nist.gov
pool.ntp.org

[ips_fixed]
r.ripple.com 51235

[validators_file]
validators.txt

[rpc_startup]
{ "command": "log_level", "severity": "warning" }

[features]
# Amendments to treat as enabled on this server. This is NOT voting:
# to vote, list amendments under [amendments] (up-vote) or
# [veto_amendments] (down-vote).
# AmendmentName

Be careful with [features]: it force-enables the listed amendments locally regardless of the network's amendment state, which can put a server out of sync with the network. Amendment voting is configured with [amendments] and [veto_amendments], which ApplicationImp::setup() feeds into the AmendmentTable.

Accessing Configuration

Components access configuration through the Application:

void SomeComponent::configure()
{
    // Get config reference
    Config const& config = app_.config();

    // Access specific sections
    auto const& nodeDB = config.section("node_db");
    auto const type = get<std::string>(nodeDB, "type");
    auto const path = get<std::string>(nodeDB, "path");

    // Access node size
    auto nodeSize = config.nodeSize;

    // Standalone mode?
    bool const standalone = config.standalone();
}

Runtime Configuration

Some settings can be adjusted at runtime via RPC:

# Change log verbosity
xrpld log_level partition severity

# Connect to peer
xrpld connect ip:port

# Get server info
xrpld server_info

Component Interaction Patterns

Pattern 1: Direct Method Calls

Most common pattern, components call each other's methods. Here is the real NetworkOPsImp::submitTransaction (src/xrpld/app/misc/NetworkOPs.cpp, trimmed) — notice it makes direct calls into other subsystems (HashRouter, LedgerMaster) to validate, but deliberately does not apply or broadcast inline; it hands the actual processing to the job queue:

void
NetworkOPsImp::submitTransaction(std::shared_ptr<STTx const> const& iTrans)
{
    // ...

    // this is an asynchronous interface
    auto const trans = sterilize(*iTrans);

    auto const txid = trans->getTransactionID();
    auto const flags = registry_.get().getHashRouter().getFlags(txid);

    // ...

    auto const [validity, reason] = checkValidity(
        registry_.get().getHashRouter(), *trans, ledgerMaster_.getValidatedRules());

    // ...

    auto tx = std::make_shared<Transaction>(trans, reason, registry_.get().getApp());

    jobQueue_.addJob(JtTransaction, "SubmitTxn", [this, tx]() {
        auto t = tx;
        processTransaction(t, false, false, FailHard::No);
    });
}

Pattern 2: Job Queue for Asynchronous Work

For work that should not block the caller. The real LedgerMaster::tryAdvance (src/xrpld/app/ledger/detail/LedgerMaster.cpp, trimmed):

void
LedgerMaster::tryAdvance()
{
    std::scoped_lock const ml(mutex_);

    // Can't advance without at least one fully-valid ledger
    advanceWork_ = true;
    if (!advanceThread_ && !validLedger_.empty())
    {
        advanceThread_ = true;
        app_.getJobQueue().addJob(JtAdvance, "AdvanceLedger", [this]() {
            // ... acquire and publish ledgers ...
        });
    }
}

Pattern 3: Event Publication

Ledger and transaction events are published to subscribers through NetworkOPs, which implements InfoSub::Source (include/xrpl/server/InfoSub.h). Subscribers (WebSocket clients, internal consumers) register through the sub* methods; the publisher side pushes events with the pub* methods:

// Subscriber side (InfoSub::Source)
virtual bool
subLedger(ref ispListener, json::Value& jvResult) = 0;
virtual bool
unsubLedger(std::uint64_t uListener) = 0;

// Publisher side (NetworkOPs)
virtual void
pubLedger(std::shared_ptr<ReadView const> const& lpAccepted) = 0;
virtual void
pubProposedTransaction(
    std::shared_ptr<ReadView const> const& ledger,
    std::shared_ptr<STTx const> const& transaction,
    TER result) = 0;
virtual void
pubValidation(std::shared_ptr<STValidation> const& val) = 0;

Pattern 4: Callback Registration

Components register callbacks for specific events, usually at construction time. Two real examples from ApplicationImp's member-initializer list (src/xrpld/app/main/Application.cpp):

// PerfLog must be started before any other threads are launched.
, perfLog_(
      perf::makePerfLog(
          perf::setupPerfLog(config_->section("perf"), config_->configDir),
          *this,
          logs_->journal("PerfLog"),
          [this] { signalStop("PerfLog"); }))

, inboundTransactions_(makeInboundTransactions(
      *this,
      collectorManager_->collector(),
      [this](std::shared_ptr<SHAMap> const& set, bool fromAcquire) {
          gotTXSet(set, fromAcquire);
      }))

Codebase Deep Dive

Key Files and Directories

Application Core:

src/xrpld/app/main/Application.h - Application interface
include/xrpl/core/ServiceRegistry.h - Subsystem accessors (Application inherits this)
src/xrpld/app/main/Application.cpp - Implementation (the ApplicationImp class lives here)
src/xrpld/app/main/Main.cpp - Entry point, creates Application

Job Queue:

include/xrpl/core/JobQueue.h - Job queue interface
src/libxrpl/core/detail/JobQueue.cpp - Implementation
include/xrpl/core/Job.h - Job definition

Configuration:

src/xrpld/core/Config.h - Config class
src/xrpld/core/detail/Config.cpp - Section name definitions

Subsystem Implementations:

Finding Application Creation

Start in Main.cpp (trimmed):

auto app =
    makeApplication(std::move(config), std::move(logs), std::make_unique<TimeKeeper>());

if (!app->setup(vm))
    return -1;

// ...

// Start the server
app->start(true /*start timers*/);

// Block until we get a stop RPC.
app->run();

return 0;

Tracing Component Access

Follow how components access each other:

// In any component
void MyComponent::work()
{
    // Access through app_
    auto& ledgerMaster = app_.getLedgerMaster();  // → ApplicationImp::getLedgerMaster()
                                                   // → return *ledgerMaster_;
}

Understanding Job Submission

Find job submissions:

# Search for addJob calls
grep -r "addJob" src/xrpld/app/

Example (from src/xrpld/app/ledger/detail/LedgerMaster.cpp):

app_.getJobQueue().addJob(JtAdvance, "AdvanceLedger", [this]() {
    // ...
});

# Submit a payment
xrpld submit '{
  "TransactionType": "Payment",
  "Account": "...",
  "Destination": "...",
  "Amount": "1000000"
}'

Watch the logs for:

JtTransaction jobs being queued (they appear in the logs under the name transaction, from JobTypes.h)
Job processing time
Queue depth changes

Step 5: Manually close a ledger

xrpld ledger_accept

Observe jobs related to ledger close:

JtAdvance (logged as advanceLedger) - Advance validated/acquired ledgers
JtPubledger (logged as publishNewLedger) - Publish a fully-accepted ledger
JtUpdatePf (logged as updatePaths) - Update pathfinding requests

Part 3: Add Custom Logging

Step 1: Modify Application.cpp

The subsystems are constructed in ApplicationImp's member-initializer list, not in the constructor body, so you can't wrap each construction with log lines. Instead, add logging at the start of the lifecycle methods (the journal member is journal_):

bool
ApplicationImp::setup(boost::program_options::variables_map const& cmdline)
{
    JLOG(journal_.info()) << "Entering setup";
    // ... existing code ...

Add a similar line at the top of start(bool withTimers) and run().

Step 2: Recompile

cd rippled/build
cmake --build . --target xrpld

Step 3: Run and observe

./xrpld --conf=xrpld.cfg --standalone

You should see your custom log messages showing the order in which the lifecycle phases run, interleaved with the existing startup logs.

Analysis Questions

Answer these based on your exploration:

What's the first subsystem created?

Why does it need to be first?

How does the job queue decide which job to process next?

What factors influence priority?

What happens if a job throws an exception?

Find the exception handling code

How many jobs are queued during a typical ledger close?

Count from your logs

What's the relationship between Application and ApplicationImp?

Why use an interface?

How would you add a new subsystem?

What's the process?
Where would you add it?

Additional Resources

Official Documentation

XRP Ledger Dev Portal: xrpl.org/docs
Rippled Setup: xrpl.org/install-rippled
Configuration Reference: xrpl.org/rippled-server-configuration

Codebase References

src/xrpld/app/main - Application layer implementation
include/xrpl/core/JobQueue.h - Job queue system
src/xrpld/core/Config.h - Configuration management
src/xrpld/app/main/Main.cpp - Program entry point

Transactors - How transactions are processed
Consensus Engine - How consensus integrates with Application
Codebase Navigation - Finding your way around the code

Summary

This module explained how the Application class holds rippled together. It is the central orchestrator that constructs and owns every subsystem, from LedgerMaster and NetworkOPs to the Overlay and NodeStore, and hands them out on request (a service-locator pattern). You saw the node's startup, start, run, and shutdown sequence, how components reach one another through the Application, and how background work is scheduled on the job queue.

To remember:

Application is the service locator: components ask it for subsystems instead of constructing their own (the accessors live on ServiceRegistry, which Application inherits)
Interface and ApplicationImp live in src/xrpld/app/main
Major subsystems: LedgerMaster, NetworkOPs, Overlay, NodeStore, JobQueue
Reach them through accessors: app.getLedgerMaster(), app.getOverlay(), app.getNodeStore()
Background work goes through the JobQueue (include/xrpl/core/JobQueue.h): app.getJobQueue().addJob(...) with a no-argument lambda, priority-scheduled (later JobType enum entries = higher priority)
The lifecycle is setup(vm) → start(true) → run(); signalStop() only sets a flag, and the teardown runs at the end of run() in an explicit, delicate order — not simply the reverse of construction
All behaviour is configured from xrpld.cfg, parsed into Config (src/xrpld/core); [features] enables amendments locally, while voting uses [amendments]/[veto_amendments]
Watch out: touching a subsystem before the Application finishes construction is the classic startup crash; respect the init order

Next up. You know who owns every subsystem; now you need to find them in over half a million lines of C++. Next: navigating the rippled codebase without getting lost.

The Application layer

Introduction

The Application Class Architecture

Design Philosophy

Application Interface

ApplicationImp Implementation

Initialization and Lifecycle

Startup Sequence

Complete Lifecycle Diagram

Subsystem Coordination

The Service Locator Pattern

Major Subsystems

Job Queue System

Purpose and Design

Job Types

Submitting Jobs

Job Priority and Scheduling

Job Queue Configuration

Configuration Management

Configuration File Structure

Accessing Configuration

Runtime Configuration

Component Interaction Patterns

Pattern 1: Direct Method Calls

Pattern 2: Job Queue for Asynchronous Work

Pattern 3: Event Publication

Pattern 4: Callback Registration

Codebase Deep Dive

Key Files and Directories

Code Navigation Tips

Additional Resources

Official Documentation

Codebase References

Summary

The Application layer

Introduction

The Application Class Architecture

Design Philosophy

Application Interface

ApplicationImp Implementation

Initialization and Lifecycle

Startup Sequence

Complete Lifecycle Diagram

Subsystem Coordination

The Service Locator Pattern

Major Subsystems

Job Queue System

Purpose and Design

Job Types

Submitting Jobs

Job Priority and Scheduling

Job Queue Configuration

Configuration Management

Configuration File Structure

Accessing Configuration

Runtime Configuration

Component Interaction Patterns

Pattern 1: Direct Method Calls

Pattern 2: Job Queue for Asynchronous Work

Pattern 3: Event Publication

Pattern 4: Callback Registration

Codebase Deep Dive

Key Files and Directories

Code Navigation Tips

Additional Resources

Official Documentation

Codebase References

Related Topics

Summary