agent/agent.go
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
// design touchstones // simple to configure/implement // - minimal configurable options // - sparse use of commands/flags/env vars // - fast to start/join a cluster // - http-only api // api auth/encryption is best implemented by a reverse proxy (like caddy or nginx) // - no RPC (it's too annoying to deal with) // predictable state // - no runtime configuration allowed (no reload support) // for a gossip cluster to work, nodes in the cluster must // be the source of truth. we do not allow the user to change // the cluster's view of reality in any way - the user may // only inspect it. // - HTTP API only supports read-only commands // this ensures that cascade's starting state never // diverges from its running state. // (restarts are very fast + interruptionless) // this also helps with security // one massive global cluster // - for simplicity, operability, and because gossip can handle it // compatability // - attempt to have a consul-compatible API where it matters? // todo // dns resolver for services // read-write exception for maintenance mode? // opentelemetry metrics // how should cascade handle name conflicts? defaulting to just blowing up // for now, but we _could_ take the tailscale route & append -1 or whatever // to the node. that would be more user friendly. // key differences from consul // - services are gossip'd // in cascade, node == member, and catalog == agent // agent api (http-only) // GET /v1/agent/metrics -> opentelemetry metrics for this agent // GET /v1/agent/members -> show serf cluster membership state // GET /v1/agent/services -> show services this agent owns // aliases: GET /v1/catalog/nodes // catalog api (http & dns) // GET /v1/catalog/nodes // GET /v1/catalog/services // dns: dig @127.0.0.1 -p 8600 nostromo.node.cascade. ANY // dns: dig @127.0.0.1 -p 8600 redis.service.cascade. ANY // consul uses /catalog because the masters maintain a service // catalog separate than the agent state // in cascade, the agents _are_ expected to maintain service catalogues, // so we do away with the catalog concept entirely. package agent import ( "fmt" "os" "sync" "time" "github.com/hashicorp/serf/serf" "golang.org/x/exp/slog" ) // Agent starts and manages a Serf & adds service discovery (TODO) type Agent struct { // this is the agent's configuration Config *Config serfConfig *serf.Config // This is the underlying Serf we are wrapping serf *serf.Serf // We receive serf events on this channel eventCh chan serf.Event logger *slog.Logger // if this channel recieves a signal, cascade shuts down shutdown bool shutdownCh chan struct{} shutdownLock sync.Mutex } // New returns an agent lmao func New(config *Config) *Agent { agent := Agent{} serfConfig := serf.DefaultConfig() eventCh := make(chan serf.Event, 1024) agent.eventCh = eventCh serfConfig.EventCh = eventCh // XXX: why do serf and cascade use the same event channel? agent.eventCh = eventCh serfConfig.EventCh = eventCh // TODO: make log level configurable // XXX: reconsider this agent-scoped logger? what does consul do? // logLogger is a bridge from the log.Logger to slog.Logger h := slog.HandlerOptions{Level: slog.LevelDebug}.NewTextHandler(os.Stderr) logger := slog.New(h) slog.SetDefault(logger) agent.logger = logger serfConfig.Logger = slog.NewLogLogger(h, slog.LevelInfo) // TODO: some of these serf settings were cargo-culted // from consul[1]. re-examine them eventually. serfConfig.EnableNameConflictResolution = false serfConfig.LeavePropagateDelay = 3 * time.Second serfConfig.MemberlistConfig.BindAddr = config.BindAddr.IP.String() serfConfig.MemberlistConfig.BindPort = config.BindAddr.Port serfConfig.MemberlistConfig.DeadNodeReclaimTime = 30 * time.Second serfConfig.MinQueueDepth = 4096 serfConfig.NodeName = config.NodeName serfConfig.ProtocolVersion = uint8(serf.ProtocolVersionMax) serfConfig.QueueDepthWarning = 1000000 serfConfig.ReconnectTimeout = 3 * 24 * time.Hour agent.Config = config agent.serfConfig = serfConfig agent.shutdownCh = make(chan struct{}) return &agent } func (a *Agent) eventLoop() { serfShutdownCh := a.serf.ShutdownCh() for { select { case e := <-a.eventCh: a.logger.Info("receive event", "event", e.String()) case <-serfShutdownCh: a.logger.Info("detect serf shutdown, turning off event loop") a.Shutdown() return case <-a.shutdownCh: return } } } // Start creates and starts an agent's serf client, eventloop, api, and more! // The agent config should not be modified after an agend has been started. func (a *Agent) Start() error { // Create serf first serf, err := serf.Create(a.serfConfig) if err != nil { return err } a.serf = serf // Start event loop go a.eventLoop() return nil } // Shutdown closes this agent and all of its processes. Should be preceded // by a Leave for a graceful shutdown. func (a *Agent) Shutdown() error { a.shutdownLock.Lock() defer a.shutdownLock.Unlock() if a.shutdown { return nil } if a.serf == nil { goto EXIT } a.logger.Info("request serf shutdown") if err := a.serf.Shutdown(); err != nil { return err } EXIT: a.logger.Info("complete serf shutdown") a.shutdown = true close(a.shutdownCh) return nil } // Join asks the Serf instance to join. See the Serf.Join function. func (a *Agent) Join(addrs []string) (n int, err error) { a.logger.Info("issue join request", "addrs", addrs) // we always ignore old events because cascade don't // care about the past n, err = a.serf.Join(addrs, true) if err != nil { return n, fmt.Errorf("Error joining: %v\n", err) } return n, err } // Leave prepares for a graceful shutdown of the agent and its processes func (a *Agent) Leave() error { if a.serf == nil { return nil } a.logger.Info("request graceful serf leave") return a.serf.Leave() } // ShutdownCh returns a channel that can be selected to wait // for the agent to perform a shutdown. func (a *Agent) ShutdownCh() <-chan struct{} { return a.shutdownCh } // [1]: sources for consul serf tweaks // https://github.com/hashicorp/consul/blob/v1.14.4/agent/consul/config.go // https://github.com/hashicorp/consul/blob/v1.14.4/lib/serf/serf.go