|
| 1 | +package batchstats |
| 2 | + |
| 3 | +import ( |
| 4 | + "context" |
| 5 | + "encoding/json" |
| 6 | + "os" |
| 7 | + "sync" |
| 8 | + "sync/atomic" |
| 9 | + "time" |
| 10 | + |
| 11 | + "github.com/google/uuid" |
| 12 | + "golang.org/x/xerrors" |
| 13 | + |
| 14 | + "cdr.dev/slog" |
| 15 | + "cdr.dev/slog/sloggers/sloghuman" |
| 16 | + |
| 17 | + "github.com/coder/coder/coderd/database" |
| 18 | + "github.com/coder/coder/coderd/database/dbauthz" |
| 19 | + "github.com/coder/coder/codersdk/agentsdk" |
| 20 | +) |
| 21 | + |
| 22 | +const ( |
| 23 | + defaultBufferSize = 1024 |
| 24 | + defaultFlushInterval = time.Second |
| 25 | +) |
| 26 | + |
| 27 | +// Batcher holds a buffer of agent stats and periodically flushes them to |
| 28 | +// its configured store. It also updates the workspace's last used time. |
| 29 | +type Batcher struct { |
| 30 | + store database.Store |
| 31 | + log slog.Logger |
| 32 | + |
| 33 | + mu sync.Mutex |
| 34 | + // TODO: make this a buffered chan instead? |
| 35 | + buf *database.InsertWorkspaceAgentStatsParams |
| 36 | + // NOTE: we batch this separately as it's a jsonb field and |
| 37 | + // pq.Array + unnest doesn't play nicely with this. |
| 38 | + connectionsByProto []map[string]int64 |
| 39 | + batchSize int |
| 40 | + |
| 41 | + // tickCh is used to periodically flush the buffer. |
| 42 | + tickCh <-chan time.Time |
| 43 | + ticker *time.Ticker |
| 44 | + interval time.Duration |
| 45 | + // flushLever is used to signal the flusher to flush the buffer immediately. |
| 46 | + flushLever chan struct{} |
| 47 | + flushForced atomic.Bool |
| 48 | + // flushed is used during testing to signal that a flush has completed. |
| 49 | + flushed chan<- int |
| 50 | +} |
| 51 | + |
| 52 | +// Option is a functional option for configuring a Batcher. |
| 53 | +type Option func(b *Batcher) |
| 54 | + |
| 55 | +// WithStore sets the store to use for storing stats. |
| 56 | +func WithStore(store database.Store) Option { |
| 57 | + return func(b *Batcher) { |
| 58 | + b.store = store |
| 59 | + } |
| 60 | +} |
| 61 | + |
| 62 | +// WithBatchSize sets the number of stats to store in a batch. |
| 63 | +func WithBatchSize(size int) Option { |
| 64 | + return func(b *Batcher) { |
| 65 | + b.batchSize = size |
| 66 | + } |
| 67 | +} |
| 68 | + |
| 69 | +// WithInterval sets the interval for flushes. |
| 70 | +func WithInterval(d time.Duration) Option { |
| 71 | + return func(b *Batcher) { |
| 72 | + b.interval = d |
| 73 | + } |
| 74 | +} |
| 75 | + |
| 76 | +// WithLogger sets the logger to use for logging. |
| 77 | +func WithLogger(log slog.Logger) Option { |
| 78 | + return func(b *Batcher) { |
| 79 | + b.log = log |
| 80 | + } |
| 81 | +} |
| 82 | + |
| 83 | +// New creates a new Batcher and starts it. |
| 84 | +func New(ctx context.Context, opts ...Option) (*Batcher, func(), error) { |
| 85 | + b := &Batcher{} |
| 86 | + b.log = slog.Make(sloghuman.Sink(os.Stderr)) |
| 87 | + b.flushLever = make(chan struct{}, 1) // Buffered so that it doesn't block. |
| 88 | + for _, opt := range opts { |
| 89 | + opt(b) |
| 90 | + } |
| 91 | + |
| 92 | + if b.store == nil { |
| 93 | + return nil, nil, xerrors.Errorf("no store configured for batcher") |
| 94 | + } |
| 95 | + |
| 96 | + if b.interval == 0 { |
| 97 | + b.interval = defaultFlushInterval |
| 98 | + } |
| 99 | + |
| 100 | + if b.batchSize == 0 { |
| 101 | + b.batchSize = defaultBufferSize |
| 102 | + } |
| 103 | + |
| 104 | + if b.tickCh == nil { |
| 105 | + b.ticker = time.NewTicker(b.interval) |
| 106 | + b.tickCh = b.ticker.C |
| 107 | + } |
| 108 | + |
| 109 | + cancelCtx, cancelFunc := context.WithCancel(ctx) |
| 110 | + done := make(chan struct{}) |
| 111 | + go func() { |
| 112 | + b.run(cancelCtx) |
| 113 | + close(done) |
| 114 | + }() |
| 115 | + |
| 116 | + closer := func() { |
| 117 | + cancelFunc() |
| 118 | + if b.ticker != nil { |
| 119 | + b.ticker.Stop() |
| 120 | + } |
| 121 | + <-done |
| 122 | + } |
| 123 | + |
| 124 | + return b, closer, nil |
| 125 | +} |
| 126 | + |
| 127 | +// Add adds a stat to the batcher for the given workspace and agent. |
| 128 | +func (b *Batcher) Add( |
| 129 | + agentID uuid.UUID, |
| 130 | + templateID uuid.UUID, |
| 131 | + userID uuid.UUID, |
| 132 | + workspaceID uuid.UUID, |
| 133 | + st agentsdk.Stats, |
| 134 | +) error { |
| 135 | + b.mu.Lock() |
| 136 | + defer b.mu.Unlock() |
| 137 | + |
| 138 | + now := database.Now() |
| 139 | + |
| 140 | + b.buf.ID = append(b.buf.ID, uuid.New()) |
| 141 | + b.buf.CreatedAt = append(b.buf.CreatedAt, now) |
| 142 | + b.buf.AgentID = append(b.buf.AgentID, agentID) |
| 143 | + b.buf.UserID = append(b.buf.UserID, userID) |
| 144 | + b.buf.TemplateID = append(b.buf.TemplateID, templateID) |
| 145 | + b.buf.WorkspaceID = append(b.buf.WorkspaceID, workspaceID) |
| 146 | + |
| 147 | + // Store the connections by proto separately as it's a jsonb field. We marshal on flush. |
| 148 | + // b.buf.ConnectionsByProto = append(b.buf.ConnectionsByProto, st.ConnectionsByProto) |
| 149 | + b.connectionsByProto = append(b.connectionsByProto, st.ConnectionsByProto) |
| 150 | + |
| 151 | + b.buf.ConnectionCount = append(b.buf.ConnectionCount, st.ConnectionCount) |
| 152 | + b.buf.RxPackets = append(b.buf.RxPackets, st.RxPackets) |
| 153 | + b.buf.RxBytes = append(b.buf.RxBytes, st.RxBytes) |
| 154 | + b.buf.TxPackets = append(b.buf.TxPackets, st.TxPackets) |
| 155 | + b.buf.TxBytes = append(b.buf.TxBytes, st.TxBytes) |
| 156 | + b.buf.SessionCountVSCode = append(b.buf.SessionCountVSCode, st.SessionCountVSCode) |
| 157 | + b.buf.SessionCountJetBrains = append(b.buf.SessionCountJetBrains, st.SessionCountJetBrains) |
| 158 | + b.buf.SessionCountReconnectingPTY = append(b.buf.SessionCountReconnectingPTY, st.SessionCountReconnectingPTY) |
| 159 | + b.buf.SessionCountSSH = append(b.buf.SessionCountSSH, st.SessionCountSSH) |
| 160 | + b.buf.ConnectionMedianLatencyMS = append(b.buf.ConnectionMedianLatencyMS, st.ConnectionMedianLatencyMS) |
| 161 | + |
| 162 | + // If the buffer is over 80% full, signal the flusher to flush immediately. |
| 163 | + // We want to trigger flushes early to reduce the likelihood of |
| 164 | + // accidentally growing the buffer over batchSize. |
| 165 | + filled := float64(len(b.buf.ID)) / float64(b.batchSize) |
| 166 | + if filled >= 0.8 && !b.flushForced.Load() { |
| 167 | + b.flushLever <- struct{}{} |
| 168 | + b.flushForced.Store(true) |
| 169 | + } |
| 170 | + return nil |
| 171 | +} |
| 172 | + |
| 173 | +// Run runs the batcher. |
| 174 | +func (b *Batcher) run(ctx context.Context) { |
| 175 | + b.initBuf(b.batchSize) |
| 176 | + // nolint:gocritic // This is only ever used for one thing - inserting agent stats. |
| 177 | + authCtx := dbauthz.AsSystemRestricted(ctx) |
| 178 | + for { |
| 179 | + select { |
| 180 | + case <-b.tickCh: |
| 181 | + b.flush(authCtx, false, "scheduled") |
| 182 | + case <-b.flushLever: |
| 183 | + // If the flush lever is depressed, flush the buffer immediately. |
| 184 | + b.flush(authCtx, true, "reaching capacity") |
| 185 | + case <-ctx.Done(): |
| 186 | + b.log.Warn(ctx, "context done, flushing before exit") |
| 187 | + b.flush(authCtx, true, "exit") |
| 188 | + return |
| 189 | + } |
| 190 | + } |
| 191 | +} |
| 192 | + |
| 193 | +// flush flushes the batcher's buffer. |
| 194 | +func (b *Batcher) flush(ctx context.Context, forced bool, reason string) { |
| 195 | + b.mu.Lock() |
| 196 | + b.flushForced.Store(true) |
| 197 | + start := time.Now() |
| 198 | + count := len(b.buf.ID) |
| 199 | + defer func() { |
| 200 | + b.flushForced.Store(false) |
| 201 | + b.mu.Unlock() |
| 202 | + // Notify that a flush has completed. This only happens in tests. |
| 203 | + if b.flushed != nil { |
| 204 | + select { |
| 205 | + case <-ctx.Done(): |
| 206 | + close(b.flushed) |
| 207 | + default: |
| 208 | + b.flushed <- count |
| 209 | + } |
| 210 | + } |
| 211 | + if count > 0 { |
| 212 | + elapsed := time.Since(start) |
| 213 | + b.log.Debug(ctx, "flush complete", |
| 214 | + slog.F("count", count), |
| 215 | + slog.F("elapsed", elapsed), |
| 216 | + slog.F("forced", forced), |
| 217 | + slog.F("reason", reason), |
| 218 | + ) |
| 219 | + } |
| 220 | + }() |
| 221 | + |
| 222 | + if len(b.buf.ID) == 0 { |
| 223 | + return |
| 224 | + } |
| 225 | + |
| 226 | + // marshal connections by proto |
| 227 | + payload, err := json.Marshal(b.connectionsByProto) |
| 228 | + if err != nil { |
| 229 | + b.log.Error(ctx, "unable to marshal agent connections by proto, dropping data", slog.Error(err)) |
| 230 | + b.buf.ConnectionsByProto = json.RawMessage(`[]`) |
| 231 | + } else { |
| 232 | + b.buf.ConnectionsByProto = payload |
| 233 | + } |
| 234 | + |
| 235 | + err = b.store.InsertWorkspaceAgentStats(ctx, *b.buf) |
| 236 | + elapsed := time.Since(start) |
| 237 | + if err != nil { |
| 238 | + b.log.Error(ctx, "error inserting workspace agent stats", slog.Error(err), slog.F("elapsed", elapsed)) |
| 239 | + return |
| 240 | + } |
| 241 | + |
| 242 | + b.resetBuf() |
| 243 | +} |
| 244 | + |
| 245 | +// initBuf resets the buffer. b MUST be locked. |
| 246 | +func (b *Batcher) initBuf(size int) { |
| 247 | + b.buf = &database.InsertWorkspaceAgentStatsParams{ |
| 248 | + ID: make([]uuid.UUID, 0, b.batchSize), |
| 249 | + CreatedAt: make([]time.Time, 0, b.batchSize), |
| 250 | + UserID: make([]uuid.UUID, 0, b.batchSize), |
| 251 | + WorkspaceID: make([]uuid.UUID, 0, b.batchSize), |
| 252 | + TemplateID: make([]uuid.UUID, 0, b.batchSize), |
| 253 | + AgentID: make([]uuid.UUID, 0, b.batchSize), |
| 254 | + ConnectionsByProto: json.RawMessage("[]"), |
| 255 | + ConnectionCount: make([]int64, 0, b.batchSize), |
| 256 | + RxPackets: make([]int64, 0, b.batchSize), |
| 257 | + RxBytes: make([]int64, 0, b.batchSize), |
| 258 | + TxPackets: make([]int64, 0, b.batchSize), |
| 259 | + TxBytes: make([]int64, 0, b.batchSize), |
| 260 | + SessionCountVSCode: make([]int64, 0, b.batchSize), |
| 261 | + SessionCountJetBrains: make([]int64, 0, b.batchSize), |
| 262 | + SessionCountReconnectingPTY: make([]int64, 0, b.batchSize), |
| 263 | + SessionCountSSH: make([]int64, 0, b.batchSize), |
| 264 | + ConnectionMedianLatencyMS: make([]float64, 0, b.batchSize), |
| 265 | + } |
| 266 | + |
| 267 | + b.connectionsByProto = make([]map[string]int64, 0, size) |
| 268 | +} |
| 269 | + |
| 270 | +func (b *Batcher) resetBuf() { |
| 271 | + b.buf.ID = b.buf.ID[:0] |
| 272 | + b.buf.CreatedAt = b.buf.CreatedAt[:0] |
| 273 | + b.buf.UserID = b.buf.UserID[:0] |
| 274 | + b.buf.WorkspaceID = b.buf.WorkspaceID[:0] |
| 275 | + b.buf.TemplateID = b.buf.TemplateID[:0] |
| 276 | + b.buf.AgentID = b.buf.AgentID[:0] |
| 277 | + b.buf.ConnectionsByProto = json.RawMessage(`[]`) |
| 278 | + b.buf.ConnectionCount = b.buf.ConnectionCount[:0] |
| 279 | + b.buf.RxPackets = b.buf.RxPackets[:0] |
| 280 | + b.buf.RxBytes = b.buf.RxBytes[:0] |
| 281 | + b.buf.TxPackets = b.buf.TxPackets[:0] |
| 282 | + b.buf.TxBytes = b.buf.TxBytes[:0] |
| 283 | + b.buf.SessionCountVSCode = b.buf.SessionCountVSCode[:0] |
| 284 | + b.buf.SessionCountJetBrains = b.buf.SessionCountJetBrains[:0] |
| 285 | + b.buf.SessionCountReconnectingPTY = b.buf.SessionCountReconnectingPTY[:0] |
| 286 | + b.buf.SessionCountSSH = b.buf.SessionCountSSH[:0] |
| 287 | + b.buf.ConnectionMedianLatencyMS = b.buf.ConnectionMedianLatencyMS[:0] |
| 288 | + b.connectionsByProto = b.connectionsByProto[:0] |
| 289 | +} |
0 commit comments