HugeGraph-PD Quick Start

1 HugeGraph-PD Overview

HugeGraph-PD (Placement Driver) is the metadata management component of HugeGraph’s distributed version, responsible for managing the distribution of graph data and coordinating storage nodes. It plays a central role in distributed HugeGraph, maintaining cluster status and coordinating HugeGraph-Store storage nodes.

PD keeps cluster metadata in an embedded RocksDB store under pd.data-path and replicates it across PD nodes with Raft, so a 3-node or 5-node PD cluster keeps serving while a minority of nodes is down. On top of that it registers and activates Store nodes, allocates and rebalances partitions, tracks Store heartbeats, and answers service discovery queries from Store and Server.

PD listens on three ports:

2 Prerequisites

2.1 Requirements

• Operating System: Linux or macOS (Windows has not been fully tested)
• Java version: ≥ 11
• Maven version: ≥ 3.5.0

3 Deployment

There are two ways to deploy the HugeGraph-PD component:

• Method 1: Download the tar package
• Method 2: Compile from source

3.1 Download the tar package

Download the latest version of HugeGraph-PD from the Apache HugeGraph official download page:

# 1.7.0 is a historical release from the incubation period, so its file and directory names still include "incubating"
wget https://downloads.apache.org/hugegraph/1.7.0/apache-hugegraph-incubating-1.7.0.tar.gz
tar zxf apache-hugegraph-incubating-1.7.0.tar.gz
cd apache-hugegraph-incubating-1.7.0/apache-hugegraph-pd-incubating-1.7.0

3.2 Compile from source

# 1. Clone the source code
git clone https://github.com/apache/hugegraph.git

# 2. Build the project
cd hugegraph
mvn clean install -DskipTests=true

# 3. After a successful build, the PD directory and packages are located at
#    hugegraph-pd/apache-hugegraph-pd-{version}          (unpacked PD distribution)
#    hugegraph-pd/apache-hugegraph-pd-{version}.tar.gz   (PD only package, Linux build hosts only)
#    target/apache-hugegraph-{version}.tar.gz            (PD + Store + Server package)

To build only the PD distribution and the modules it depends on:

mvn clean package -pl hugegraph-pd/hg-pd-dist -am -DskipTests

The unpacked distribution contains just three directories: bin (start and stop scripts), conf (application.yml, application.yml.template, log4j2.xml, verify-license.json) and lib (the hg-pd-service jar).

3.3 Docker Deployment

The HugeGraph-PD Docker image is available on Docker Hub as hugegraph/pd.

Note: The following steps assume you have already cloned or pulled the HugeGraph main repository locally, or at least have its docker/ directory available.

Use the docker compose setup to deploy the complete 3-node cluster (PD + Store + Server):

cd hugegraph/docker
# Keep the version aligned with the latest release, for example 1.x.0
HUGEGRAPH_VERSION=1.7.0 docker compose -f docker-compose-3pd-3store-3server.yml up -d

A single PD plus a single Store and Server is also available as docker-compose-hstore.yml.

To run a single PD node via docker run, configuration is provided via environment variables:

docker run -d \
  -p 8620:8620 \
  -p 8686:8686 \
  -p 8610:8610 \
  -e HG_PD_GRPC_HOST=<your-ip> \
  -e HG_PD_RAFT_ADDRESS=<your-ip>:8610 \
  -e HG_PD_RAFT_PEERS_LIST=<your-ip>:8610 \
  -e HG_PD_INITIAL_STORE_LIST=<store-ip>:8500 \
  -v /path/to/data:/hugegraph-pd/pd_data \
  --name hugegraph-pd \
  hugegraph/pd:1.7.0

Environment variable reference:

The entrypoint refuses to start if any of the four required variables is missing, and it turns the values above into a SPRING_APPLICATION_JSON override, so the packaged conf/application.yml does not need editing. Any key not covered by an HG_PD_* variable keeps the value from that file. JAVA_OPTS is passed through to the JVM.

Note: In Docker bridge networking, use container hostnames (e.g. pd0) for HG_PD_GRPC_HOST and HG_PD_RAFT_ADDRESS instead of IP addresses.

Deprecated aliases: GRPC_HOST, RAFT_ADDRESS, RAFT_PEERS, PD_INITIAL_STORE_LIST still work but log a deprecation warning. Use the HG_PD_* names for new deployments.

The image ships a HEALTHCHECK that polls GET /v1/health on port 8620 every 15 seconds, with a 90 second start period and 3 retries, so docker ps reports real PD health. The entrypoint runs the start script with -d false, so the container process is Java itself and Docker’s restart policy fires when it dies. The image also sets STDOUT_MODE=true, so docker logs <container-name> (e.g. docker logs hg-pd0) shows the PD log without exec-ing into the container.

See docker/README.md for the full cluster setup guide.

4 Configuration

The main configuration file for PD is conf/application.yml. This is the file the distribution ships:

spring:
  application:
    name: hugegraph-pd

management:
  metrics:
    export:
      prometheus:
        enabled: true
  endpoints:
    web:
      exposure:
        include: "*"

logging:
  config: 'file:./conf/log4j2.xml'

license:
  verify-path: ./conf/verify-license.json
  license-path: ./conf/hugegraph.license

grpc:
  # gRPC port for cluster mode
  port: 8686
  # Change to the actual local IPv4 address when deploying
  host: 127.0.0.1

server:
  # REST service port
  port: 8620

pd:
  # Storage path
  data-path: ./pd_data
  # Auto-expansion check cycle (seconds)
  patrol-interval: 1800
  # Minimum number of Store nodes required for cluster availability
  initial-store-count: 1
  # Store configuration information, format is IP:gRPC port
  initial-store-list: 127.0.0.1:8500

raft:
  # Raft address of this node
  address: 127.0.0.1:8610
  # Raft addresses of all PD nodes in the cluster
  peers-list: 127.0.0.1:8610

store:
  # Store offline time (seconds). After this time, the store is considered permanently unavailable
  max-down-time: 172800
  # Whether to enable store monitoring data storage
  monitor_data_enabled: true
  # Monitoring data interval
  monitor_data_interval: 1 minute
  # Monitoring data retention time
  monitor_data_retention: 1 day

partition:
  # Default number of replicas per partition
  default-shard-count: 1
  # Default maximum number of replicas per machine
  store-max-shard-count: 12

conf/application.yml.template is a second, unused copy with placeholders ($GRPC_PORT$, $RAFT_ADDRESS$ and so on) for deployment tooling that generates the file. PD always reads conf/application.yml, which the start script passes as -Dspring.config.location.

4.1 Configuration reference

Keys not present in conf/application.yml fall back to the built-in default listed below. Keys with no built-in default must be present, otherwise PD fails to start.

gRPC and REST

application.yml.template also carries grpc.netty-server.max-inbound-message-size: 100MB, but PD sets the gRPC server’s inbound message limit to 1 GB in code, so that key has no effect.

Raft

PD core

Store management

Partitions

The initial partition count is derived from these two values and the size of pd.initial-store-list:

initial partitions = store count * partition.store-max-shard-count / partition.default-shard-count

Discovery, license and metrics

Thread pools

4.2 Single-node configuration

The shipped conf/application.yml already is a working single-node configuration. It is meant for development and testing: one PD node has no Raft quorum to lose, and partition.default-shard-count: 1 keeps a single replica per partition.

grpc:
  host: 127.0.0.1
  port: 8686
server:
  port: 8620
raft:
  address: 127.0.0.1:8610
  peers-list: 127.0.0.1:8610
pd:
  data-path: ./pd_data
  initial-store-count: 1
  initial-store-list: 127.0.0.1:8500
partition:
  default-shard-count: 1

4.3 Three-node cluster configuration

For a production cluster run 3 or 5 PD nodes, an odd number so Raft always has a quorum. A 3-node cluster tolerates one node failure. raft.peers-list must list every node and must be byte-for-byte identical on all of them, while grpc.host and raft.address differ per node.

Node 1 (192.168.1.10):

grpc:
  host: 192.168.1.10
  port: 8686
server:
  port: 8620
raft:
  address: 192.168.1.10:8610
  peers-list: 192.168.1.10:8610,192.168.1.11:8610,192.168.1.12:8610
pd:
  data-path: /data/pd
  initial-store-count: 3
  initial-store-list: 192.168.1.20:8500,192.168.1.21:8500,192.168.1.22:8500
partition:
  default-shard-count: 3

Node 2 (192.168.1.11) and node 3 (192.168.1.12) use the same file with grpc.host and raft.address changed to their own address:

# node 2
grpc:
  host: 192.168.1.11
raft:
  address: 192.168.1.11:8610
  peers-list: 192.168.1.10:8610,192.168.1.11:8610,192.168.1.12:8610

# node 3
grpc:
  host: 192.168.1.12
raft:
  address: 192.168.1.12:8610
  peers-list: 192.168.1.10:8610,192.168.1.11:8610,192.168.1.12:8610

To put all three PD nodes on one machine for testing, give each node its own pd.data-path and its own ports, for example raft 8610/8611/8612, gRPC 8686/8687/8688 and REST 8620/8621/8622.

In Docker bridge networking the same configuration comes from environment variables and uses container hostnames instead of IP addresses:

# pd0
HG_PD_GRPC_HOST: pd0
HG_PD_RAFT_ADDRESS: pd0:8610
HG_PD_RAFT_PEERS_LIST: pd0:8610,pd1:8610,pd2:8610
HG_PD_INITIAL_STORE_LIST: store0:8500,store1:8500,store2:8500
HG_PD_INITIAL_STORE_COUNT: 3

# pd1
HG_PD_GRPC_HOST: pd1
HG_PD_RAFT_ADDRESS: pd1:8610
HG_PD_RAFT_PEERS_LIST: pd0:8610,pd1:8610,pd2:8610

# pd2
HG_PD_GRPC_HOST: pd2
HG_PD_RAFT_ADDRESS: pd2:8610
HG_PD_RAFT_PEERS_LIST: pd0:8610,pd1:8610,pd2:8610

5 Start and Stop

5.1 Start PD

In the PD installation directory, execute:

./bin/start-hugegraph-pd.sh

The script requires a JDK of at least version 11 on PATH or in JAVA_HOME, and it exits without doing anything if it finds a Java process already using this installation’s conf directory.

Supported flags:

The -d flag controls daemon mode:

• -d true (default): run as a background daemon; the script returns immediately.
• -d false: run in foreground. The script execs Java, so the container or supervisor process IS Java. Use this when running under Docker or a process supervisor (systemd, supervisord) so crashes are detected and the service is restarted automatically.

Each flag also has an environment variable equivalent: DAEMON, GC_OPTION, USER_OPTION and OPEN_TELEMETRY. Setting JAVA_OPTIONS replaces the computed heap settings entirely; otherwise the script sizes the heap between 512 MB and 32 GB from available memory. Setting STDOUT_MODE=true leaves the JVM output on stdout instead of redirecting it to logs/hugegraph-pd-stdout.log, which is what the Docker image does.

After successful startup, you can see logs similar to the following in logs/hugegraph-pd-stdout.log:

YYYY-mm-dd xx:xx:xx [main] [INFO] o.a.h.p.b.HugePDServer - Started HugePDServer in x.xxx seconds (JVM running for x.xxx)

The process id is written to bin/pid.

5.2 Stop PD

In the PD installation directory, execute:

./bin/stop-hugegraph-pd.sh

The script reads bin/pid, sends the process a termination signal, waits up to 30 seconds for it to exit, and removes the pid file. If bin/pid is missing it reports that and exits successfully.

6 Startup Order in a Distributed Cluster

Start the components in this order:

1. All PD nodes. They form the Raft group and elect a leader. Wait until every node answers GET /v1/health.
2. All Store nodes. Each Store registers with PD over gRPC, and PD activates the ones listed in pd.initial-store-list. Wait until GET /v1/stores reports "state": "Up" for every Store.
3. All Server nodes. A Server reads pd.peers and depends on PD reporting at least one live Store before partitions can be assigned.

The Docker Compose topologies enforce exactly this. Store containers wait on PD’s /v1/health healthcheck through depends_on with condition: service_healthy, Server containers wait the same way on the Store healthcheck, and the Server entrypoint then polls PD’s /v1/stores until a Store reports Up before it starts HugeGraph.

PD is also the last component to stop: shut down Server, then Store, then PD.

7 Verification

7.1 REST API authentication

Except for /actuator/*, /v1/health and /v1/prom/targets/*, every PD REST path requires an HTTP Basic Authorization header whose user name is one of the internal service names hg, store, hubble or vermeer. A request without the header is answered with:

{"status": -1, "error": "Unauthorized!"}

The password is not validated yet, so any value works. The Server’s own bin/wait-storage.sh uses store:admin and lets you override it with PD_AUTH_USER and PD_AUTH_PASSWORD, so the examples below use the same credentials:

curl -u store:admin http://localhost:8620/v1/stores

Warning: This check is only meant to separate HugeGraph’s own components from other traffic. Do not expose the PD REST or gRPC ports to an untrusted network. Restrict them with firewall rules or security groups, and keep raft.ip-whitelist.enabled on so the Raft port only accepts the configured peers.

7.2 Health check

GET /v1/health needs no credentials and is what the Docker healthcheck uses. It answers 200 with an empty body:

curl -i http://localhost:8620/v1/health

The Spring Boot actuator endpoint also works and is more readable:

curl http://localhost:8620/actuator/health

If it returns {"status":"UP"}, it indicates that the PD service has been successfully started.

7.3 Cluster and member status

Check the PD members and which node is the Raft leader:

curl -u store:admin http://localhost:8620/v1/members

The response carries pdList, the elected pdLeader, numOfService, numOfNormalService and a stateCountMap. In a healthy 3-node PD cluster numOfService and numOfNormalService are both 3 and exactly one member has role: "Leader".

GET /v1/cluster returns the same member list together with the Store list, graph list and overall cluster state, and GET / returns a short summary (leader address, cluster state, member count, store count, graph count, partition count).

7.4 Store status

You can also verify Store node status through the PD API:

curl -u store:admin http://localhost:8620/v1/stores

If the response shows state as Up, the corresponding Store node is running normally. The example below shows a single Store node. In a healthy 3-node deployment, the storeId list should contain three IDs, and stateCountMap.Up, numOfService, and numOfNormalService should all be 3.

{
  "message": "OK",
  "data": {
    "stores": [
      {
        "storeId": 8319292642220586694,
        "address": "127.0.0.1:8500",
        "raftAddress": "127.0.0.1:8510",
        "version": "",
        "state": "Up",
        "deployPath": "/Users/{your_user_name}/hugegraph/apache-hugegraph-incubating-1.7.0/apache-hugegraph-store-incubating-1.7.0/lib/hg-store-node-1.7.0.jar",
        "dataPath": "./storage",
        "startTimeStamp": 1754027127969,
        "registedTimeStamp": 1754027127969,
        "lastHeartBeat": 1754027909444,
        "capacity": 494384795648,
        "available": 346535829504,
        "partitionCount": 0,
        "graphSize": 0,
        "keyCount": 0,
        "leaderCount": 0,
        "serviceName": "127.0.0.1:8500-store",
        "serviceVersion": "",
        "serviceCreatedTimeStamp": 1754027127000,
        "partitions": []
      }
    ],
    "stateCountMap": {
      "Up": 1
    },
    "numOfService": 1,
    "numOfNormalService": 1
  },
  "status": 0
}

7.5 Other REST endpoints

All paths below are relative to http://<pd-host>:8620 and need the Basic header from section 7.1 unless noted.

The two log endpoints take a time range as {"startTime": "...", "endTime": "..."}; yyyy-MM-dd HH:mm:ss and yyyy-MM-dd are among the accepted formats.

PD registers its own meters under the hg prefix, so /actuator/prometheus exposes hg_up, hg_graphs, hg_stores and hg_terms alongside the standard JVM metrics, plus per-graph partition and size meters once graphs exist.