現在テスト中です: 完了後にGitHubのコードを公開する予定です。

設定 (環境変数)

設定は環境変数だけから読みます。nameは大文字小文字を区別せず、空値は未設定、1/true/yes/onだけがtrueです。不正な範囲/組合せはstartupを中止します。scratch imageはlocal disk/PVを使いません。

共通設定とresource上限

VariableDefaultAllowedPurpose
SERVER_PORT5225valid u16 portsingle client port; cluster advertises CLUSTER_SELF
SOCKET_BIND0.0.0.0bindable addressclient and Raft bind host
DEBUG_MODEbuild settingflagdebug logging
DOCKER_MODEfalseflagofficial image binds internal 5225/6225
MAX_CONNECTIONS10241..8192concurrent client connections
MAX_RAFT_CONNECTIONS641..256separate Raft/control reserve
MAX_WAITERS20481..16384waiters per key; excess acquire gets B
MAX_TOTAL_WAITERS16384MAX_WAITERS..65536total waiters; excess acquire gets B
Hard capValueSaturation behavior
in-flight acquire per connection64B for that acquire
global in-flight acquire4096B for that acquire
global in-flight release512 separate lanebounded wait until connection close
queued replies per connection256close slow client
coordinator channel4096backpressure
cluster active keys65536B for new-key acquire
Raft frame64MiBreject RPC
encoded snapshot64MiB - 64KiBreject build/install
concurrent Raft decode budget72MiBbounded permit wait/timeout

default profileは512MiB memory limit向けです。hard ceilingを超えられず、unlimited設定はありません。

client認証とTLS

VariableDefaultPurpose
CLIENT_TOKENSone empty tokencomma-separated client credentials; production should use nonblank secrets
TLS_CERT / TLS_KEYunsetserver certificate chain/private key; set together
TLS_CAsystem rootsCA for peer certificate verification
TLS_SKIP_VERIFYfalsetest only
CLUSTER_ALLOW_PLAINTEXT_PRIVATEfalseexplicit opt-out for isolated private networks

TLSを推奨します。plaintext clusterはprivate network向け明示opt-outが必要で、機密性やcertificate-based node identityを提供しません。

安定したvoter identity

VariableConstraintPurpose
CLUSTER_ID1–128 ASCII alnum/./_/-fencing domain
CLUSTER_NODE_IDu64Raft ID; never reuse after retirement
CLUSTER_INCARNATIONnon-zero u128freshly issued per process start
CLUSTER_SELFhost:client-portstable client slot and redirect address
CLUSTER_RAFT_SELFhost:raft-portunique Raft/control address for this process
CLUSTER_PEERSexactly 3 or 5NodeId@incarnation@client-address@raft-address entries

cluster identity変数は全て一緒に必須で、全省略はsingle modeです。process startごとに外部発行の新incarnationが必要です。retired NodeIdは再利用せず、replacementは新NodeId learnerです。client/Raft portは明示し、+1000規則はありません。

startup admissionはfail-closedです。初期clusterは全peerがempty/unconsumedを証明しPrepared barrierを通って初めてOpenです。replacementはone-shot Join承認後のみOpenです。Pending/PreparedはVote/Append/Snapshot/client mutationを拒否し、unreachable peerをemptyと推定しません。

Raftと認証clock

VariableDefaultConstraint
CLUSTER_HEARTBEAT_MS500>=10
CLUSTER_ELECTION_TIMEOUT_MS2000>=600, >=4× heartbeat
CLUSTER_SNAPSHOT_TIMEOUT_MS120000>=1000
RAFT_PROGRESS_TIMEOUT_MS10000>=1000
CLUSTER_CLOCK_AGENT_ENDPOINTunsetrequired numeric IP:port in cluster
CLUSTER_CLOCK_AGENT_SECRETunsetrequired, at least 32 bytes, separate domain
CLUSTER_CLOCK_PAIRWISE_DELTA_MS202..=60000, identical on all nodes
CLUSTER_CLOCK_MAX_STALENESS_MS5000greater than sample interval + timeout
CLUSTER_CLOCK_SAMPLE_INTERVAL_MS1000positive and below staleness
CLUSTER_CLOCK_TIMEOUT_MS250positive and below staleness
CLUSTER_CLOCK_MAX_RTT_MS50positive and at most timeout
CLUSTER_CLOCK_MAX_UNCERTAINTY_MS2at most pairwise delta / 2
CLUSTER_CLOCK_MAX_SOURCE_AGE_MS2000positive

clock endpointは認証済みreference time、uncertainty、sync state、source ageを返します。auth/freshness/RTT/pairwise delta/wall-step検査失敗は新mutationを停止しclientをdrainしてRaftを終了します。

cluster認証とreplacement join

VariableRequirement
CLUSTER_TOKENSnonblank list required in cluster
CLUSTER_CONTROL_TOKENSnonblank list required in cluster
CLUSTER_BOOTSTRAP_CREDENTIALinitial cluster process only; recorded once in replicated state

Replacement process values generated by prepare-replacement:

  • CLUSTER_JOIN_LEADER
  • CLUSTER_JOIN_LEADER_NODE_ID
  • CLUSTER_JOIN_LEADER_INCARNATION
  • CLUSTER_JOIN_CREDENTIAL
  • CLUSTER_JOIN_SEED_PEERS
  • fresh CLUSTER_NODE_ID, CLUSTER_INCARNATION, CLUSTER_SELF, CLUSTER_RAFT_SELF, CLUSTER_PEERS

cluster/control/clock/bootstrap/join credentialは別domainで再利用禁止です。Join値はprepare-replacement出力だけを使い、credentialはone-shot(default 60秒、5–300秒)です。peer wireはCLUSTER_WIRE_VERSION=1 exact matchでpackage versionは使いません。

See zero-downtime fresh-NodeId learner replacement.