| 1 | package metrics |
| 2 | |
| 3 | import ( |
| 4 | "strconv" |
| 5 | "time" |
| 6 | ) |
| 7 | |
| 8 | // Rules is the thresholds and sustain windows Evaluate checks each tick |
| 9 | // against. Frac fields are fractions (0..1), matching HostSample; LagSecs is |
| 10 | // the one exception — keymail relay lag arrives in seconds, not a fraction. |
| 11 | type Rules struct { |
| 12 | DiskFrac float64 // fire >, default 0.90, immediate |
| 13 | MemFrac float64 // default 0.90 |
| 14 | MemSustain time.Duration // default 10 * time.Minute |
| 15 | CPUFrac float64 // default 0.95 |
| 16 | CPUSustain time.Duration // default 15 * time.Minute |
| 17 | LagSecs int64 // default 900, immediate; inert while readings are nil |
| 18 | ClearBy float64 // hysteresis, default 0.05 (fractions); lag clears at LagSecs-300 |
| 19 | // GeoIPSustain is how long a geoip reason must hold before the rule |
| 20 | // fires (geoip spec §10.4); it clears on the first reading of 0. |
| 21 | GeoIPSustain time.Duration // default 30 * time.Minute |
| 22 | } |
| 23 | |
| 24 | // DefaultRules are the thresholds shipped with the platform. |
| 25 | func DefaultRules() Rules { |
| 26 | return Rules{ |
| 27 | DiskFrac: 0.90, |
| 28 | MemFrac: 0.90, |
| 29 | MemSustain: 10 * time.Minute, |
| 30 | CPUFrac: 0.95, |
| 31 | CPUSustain: 15 * time.Minute, |
| 32 | LagSecs: 900, |
| 33 | ClearBy: 0.05, |
| 34 | // GeoIPSustain: geoip spec §10.4. |
| 35 | GeoIPSustain: 30 * time.Minute, |
| 36 | } |
| 37 | } |
| 38 | |
| 39 | // RulesFromEnv applies percent-integer overrides on top of DefaultRules: |
| 40 | // CARLOS_ALERT_DISK_PCT, CARLOS_ALERT_MEM_PCT, CARLOS_ALERT_CPU_PCT. An |
| 41 | // unset, unparsable, or out-of-range (0..100) value keeps the default — |
| 42 | // deliberately pure (no logging); a caller that wants to log a rejected |
| 43 | // override asks PctFrac the same question this does, since this package |
| 44 | // touches neither stderr nor the OS. |
| 45 | func RulesFromEnv(getenv func(string) string) Rules { |
| 46 | r := DefaultRules() |
| 47 | if v, ok := PctFrac(getenv("CARLOS_ALERT_DISK_PCT")); ok { |
| 48 | r.DiskFrac = v |
| 49 | } |
| 50 | if v, ok := PctFrac(getenv("CARLOS_ALERT_MEM_PCT")); ok { |
| 51 | r.MemFrac = v |
| 52 | } |
| 53 | if v, ok := PctFrac(getenv("CARLOS_ALERT_CPU_PCT")); ok { |
| 54 | r.CPUFrac = v |
| 55 | } |
| 56 | return r |
| 57 | } |
| 58 | |
| 59 | // PctFrac converts one percent-integer override string ("95") to the |
| 60 | // fraction Rules holds (0.95), reporting ok=false for empty, unparsable, or |
| 61 | // out-of-range (0..100) input. Exported so the agent can LOG an override it |
| 62 | // set but this rejected without either duplicating the rule or dragging |
| 63 | // stderr into this package. |
| 64 | func PctFrac(s string) (float64, bool) { |
| 65 | if s == "" { |
| 66 | return 0, false |
| 67 | } |
| 68 | n, err := strconv.Atoi(s) |
| 69 | if err != nil || n < 0 || n > 100 { |
| 70 | return 0, false |
| 71 | } |
| 72 | return float64(n) / 100.0, true |
| 73 | } |
| 74 | |
| 75 | // Observation is one tick's readings for a box. A nil field means "no |
| 76 | // data this tick" and is handled identically everywhere in Evaluate: the |
| 77 | // corresponding rule's state is left untouched (a firing alert stays |
| 78 | // firing, a quiet rule stays quiet, nothing is ever emitted for it). |
| 79 | type Observation struct { |
| 80 | At time.Time |
| 81 | CPUFrac *float64 |
| 82 | MemFrac *float64 |
| 83 | DiskFrac *float64 |
| 84 | LagSecs *int64 |
| 85 | // GeoIP is the box's geoip reason code (geoip spec §10.4): 0 none, |
| 86 | // 1 stale, 2 expiring, 3 expired, 4 not loaded, 5 feed failing. nil on |
| 87 | // a box whose edge has not published one. |
| 88 | GeoIP *float64 |
| 89 | } |
| 90 | |
| 91 | // RuleState is one rule's sustain/hysteresis bookkeeping. Since and FiredAt |
| 92 | // carry `omitempty`, but encoding/json's omitempty never treats a struct |
| 93 | // (time.Time is a struct) as empty regardless of its zero value — so a |
| 94 | // never-triggered RuleState does NOT drop these keys. It marshals as |
| 95 | // {"firing":false,"since":"0001-01-01T00:00:00Z","fired_at":"0001-01-01T00:00:00Z"}, |
| 96 | // both fields always present. The round-trip is still lossless (the zero |
| 97 | // time reproduces exactly on Unmarshal), so this is a documentation note, |
| 98 | // not a bug: any consumer of this wire shape (the agent loop's persistence, |
| 99 | // the console watcher) must treat the zero time value itself — not key |
| 100 | // absence — as "never set". See TestRuleStateJSONRoundTrip. |
| 101 | type RuleState struct { |
| 102 | Firing bool `json:"firing"` |
| 103 | Since time.Time `json:"since,omitempty"` // continuously-above-threshold since (sustain tracking) |
| 104 | FiredAt time.Time `json:"fired_at,omitempty"` // set on fire, cleared on clear |
| 105 | Value float64 `json:"value,omitempty"` // reading at last transition |
| 106 | } |
| 107 | |
| 108 | // AlertState is the JSON persisted to disk by the agent loop and published |
| 109 | // to alerts/boxes/<box>.json (see AlertKey), and read back by the console |
| 110 | // watcher that diffs Rules[...].Firing across polls. |
| 111 | type AlertState struct { |
| 112 | V int `json:"v"` // 1 |
| 113 | Box string `json:"box"` |
| 114 | UpdatedAt time.Time `json:"updated_at"` |
| 115 | Rules map[string]RuleState `json:"rules"` // "disk","mem","cpu","lag" |
| 116 | } |
| 117 | |
| 118 | // Event is emitted on a firing or clearing transition — never on every |
| 119 | // tick, and never for a nil reading. |
| 120 | type Event struct { |
| 121 | Rule string `json:"rule"` |
| 122 | Firing bool `json:"firing"` // true=fired, false=cleared |
| 123 | Value float64 `json:"value"` |
| 124 | Box string `json:"box"` |
| 125 | } |
| 126 | |
| 127 | // AlertKey is the bucket key an AlertState for box is stored/published |
| 128 | // under. |
| 129 | func AlertKey(box string) string { |
| 130 | return "alerts/boxes/" + box + ".json" |
| 131 | } |
| 132 | |
| 133 | // Evaluate advances AlertState by one Observation and returns the new |
| 134 | // state plus any fire/clear transitions as Events. |
| 135 | // |
| 136 | // Evaluate never mutates st: it builds a fresh Rules map (copying each |
| 137 | // RuleState by value — RuleState holds no pointers or slices, so the copy |
| 138 | // is a true deep copy) and returns that in the new AlertState. Callers |
| 139 | // comparing old vs new state across the call — the console watcher this |
| 140 | // feeds diffs Rules[...].Firing on exactly that assumption — see two |
| 141 | // independent maps, never the same one mutated in place. |
| 142 | // |
| 143 | // Semantics, per rule ("disk", "mem", "cpu", "lag"): |
| 144 | // - nil reading: that rule's state is left exactly as-is. No data never |
| 145 | // fires and never clears — clearing needs evidence. |
| 146 | // - Not firing, reading > threshold: Since is set if it was zero; the |
| 147 | // rule fires (emitting an Event) once obs.At.Sub(Since) >= sustain |
| 148 | // (immediate rules use sustain 0, so they fire the same tick Since is |
| 149 | // set). A reading back at-or-below threshold before sustain elapses |
| 150 | // resets Since to zero — the sustain window must be continuous. |
| 151 | // - Firing, reading < threshold-ClearBy: the rule clears (emitting an |
| 152 | // Event) and Since/FiredAt are zeroed. |
| 153 | // - Firing, reading in [threshold-ClearBy, threshold] (or still above |
| 154 | // threshold): stays firing — this is the hysteresis band, and it is |
| 155 | // deliberately wider than just "below threshold" so a reading |
| 156 | // oscillating right at the line doesn't flap. |
| 157 | func Evaluate(r Rules, st AlertState, obs Observation) (AlertState, []Event) { |
| 158 | out := AlertState{ |
| 159 | V: st.V, |
| 160 | Box: st.Box, |
| 161 | UpdatedAt: obs.At, |
| 162 | Rules: make(map[string]RuleState, len(st.Rules)), |
| 163 | } |
| 164 | for name, rs := range st.Rules { |
| 165 | out.Rules[name] = rs |
| 166 | } |
| 167 | |
| 168 | var events []Event |
| 169 | emit := func(name string, firing bool, value float64) { |
| 170 | events = append(events, Event{Rule: name, Firing: firing, Value: value, Box: st.Box}) |
| 171 | } |
| 172 | |
| 173 | evalFrac := func(name string, reading *float64, threshold float64, sustain time.Duration) { |
| 174 | if reading == nil { |
| 175 | return |
| 176 | } |
| 177 | val := *reading |
| 178 | cur := out.Rules[name] |
| 179 | clearLine := threshold - r.ClearBy |
| 180 | |
| 181 | if cur.Firing { |
| 182 | if val < clearLine { |
| 183 | cur.Firing = false |
| 184 | cur.Since = time.Time{} |
| 185 | cur.FiredAt = time.Time{} |
| 186 | cur.Value = val |
| 187 | out.Rules[name] = cur |
| 188 | emit(name, false, val) |
| 189 | } |
| 190 | return |
| 191 | } |
| 192 | |
| 193 | if val > threshold { |
| 194 | if cur.Since.IsZero() { |
| 195 | cur.Since = obs.At |
| 196 | } |
| 197 | if obs.At.Sub(cur.Since) >= sustain { |
| 198 | cur.Firing = true |
| 199 | cur.FiredAt = obs.At |
| 200 | cur.Value = val |
| 201 | out.Rules[name] = cur |
| 202 | emit(name, true, val) |
| 203 | return |
| 204 | } |
| 205 | out.Rules[name] = cur |
| 206 | return |
| 207 | } |
| 208 | |
| 209 | if !cur.Since.IsZero() { |
| 210 | cur.Since = time.Time{} |
| 211 | out.Rules[name] = cur |
| 212 | } |
| 213 | } |
| 214 | |
| 215 | evalLagSecs := func(name string, reading *int64, threshold int64, clearBy int64, sustain time.Duration) { |
| 216 | if reading == nil { |
| 217 | return |
| 218 | } |
| 219 | val := float64(*reading) |
| 220 | cur := out.Rules[name] |
| 221 | clearLine := float64(threshold - clearBy) |
| 222 | thresholdF := float64(threshold) |
| 223 | |
| 224 | if cur.Firing { |
| 225 | if val < clearLine { |
| 226 | cur.Firing = false |
| 227 | cur.Since = time.Time{} |
| 228 | cur.FiredAt = time.Time{} |
| 229 | cur.Value = val |
| 230 | out.Rules[name] = cur |
| 231 | emit(name, false, val) |
| 232 | } |
| 233 | return |
| 234 | } |
| 235 | |
| 236 | if val > thresholdF { |
| 237 | if cur.Since.IsZero() { |
| 238 | cur.Since = obs.At |
| 239 | } |
| 240 | if obs.At.Sub(cur.Since) >= sustain { |
| 241 | cur.Firing = true |
| 242 | cur.FiredAt = obs.At |
| 243 | cur.Value = val |
| 244 | out.Rules[name] = cur |
| 245 | emit(name, true, val) |
| 246 | return |
| 247 | } |
| 248 | out.Rules[name] = cur |
| 249 | return |
| 250 | } |
| 251 | |
| 252 | if !cur.Since.IsZero() { |
| 253 | cur.Since = time.Time{} |
| 254 | out.Rules[name] = cur |
| 255 | } |
| 256 | } |
| 257 | |
| 258 | evalFrac("disk", obs.DiskFrac, r.DiskFrac, 0) |
| 259 | evalFrac("mem", obs.MemFrac, r.MemFrac, r.MemSustain) |
| 260 | evalFrac("cpu", obs.CPUFrac, r.CPUFrac, r.CPUSustain) |
| 261 | evalLagSecs("lag", obs.LagSecs, r.LagSecs, 300, 0) |
| 262 | // geoip: a reason code against 0.5, so any reason fires after the |
| 263 | // sustain and a reading of 0 clears (geoip spec §10.4). |
| 264 | evalFrac("geoip", obs.GeoIP, 0.5, r.GeoIPSustain) |
| 265 | // A reason that changes WHILE firing (expiring becomes expired) is a |
| 266 | // new firing event with the new code: the alert stays up, its sustain |
| 267 | // is not restarted, and the operator is told what it now means (code |
| 268 | // review R2-5). |
| 269 | if obs.GeoIP != nil { |
| 270 | if cur := out.Rules["geoip"]; cur.Firing && *obs.GeoIP > 0.5 && cur.Value != *obs.GeoIP { |
| 271 | cur.Value = *obs.GeoIP |
| 272 | out.Rules["geoip"] = cur |
| 273 | emit("geoip", true, *obs.GeoIP) |
| 274 | } |
| 275 | } |
| 276 | |
| 277 | return out, events |
| 278 | } |
| 279 | |
amadan