(requestID string)
| 139 | } |
| 140 | |
| 141 | func (e *MemoryEngine) Snapshot(requestID string) types.SLOSnapshot { |
| 142 | e.mu.Lock() |
| 143 | defer e.mu.Unlock() |
| 144 | |
| 145 | rec, ok := e.records[requestID] |
| 146 | if !ok { |
| 147 | return types.SLOSnapshot{} |
| 148 | } |
| 149 | |
| 150 | var ttft int64 |
| 151 | if !rec.startTime.IsZero() && !rec.firstTokenTime.IsZero() { |
| 152 | ttft = rec.firstTokenTime.Sub(rec.startTime).Milliseconds() |
| 153 | if ttft < 0 { |
| 154 | ttft = 0 |
| 155 | } |
| 156 | } |
| 157 | |
| 158 | var completion int64 |
| 159 | var tpot int64 |
| 160 | if !rec.startTime.IsZero() && !rec.endTime.IsZero() { |
| 161 | completion = rec.endTime.Sub(rec.startTime).Milliseconds() |
| 162 | if completion < 0 { |
| 163 | completion = 0 |
| 164 | } |
| 165 | } |
| 166 | if !rec.firstTokenTime.IsZero() && !rec.endTime.IsZero() { |
| 167 | postFirstToken := rec.endTime.Sub(rec.firstTokenTime).Milliseconds() |
| 168 | if postFirstToken > 0 { |
| 169 | tpot = postFirstToken / 32 |
| 170 | if tpot <= 0 { |
| 171 | tpot = 1 |
| 172 | } |
| 173 | } |
| 174 | } |
| 175 | |
| 176 | return types.SLOSnapshot{ |
| 177 | TTFTMs: ttft, |
| 178 | TPOTMs: tpot, |
| 179 | CompletionLatencyMs: completion, |
| 180 | FallbackTriggered: rec.fallback, |
| 181 | } |
| 182 | } |
| 183 | |
| 184 | func (e *MemoryEngine) ensureRecord(requestID string) *requestTiming { |
| 185 | rec, ok := e.records[requestID] |
no outgoing calls