157 lines
4.6 KiB
Go
157 lines
4.6 KiB
Go
package service
|
|
|
|
import (
|
|
edgeevents "iop/apps/edge/internal/events"
|
|
eventpkg "iop/packages/go/events"
|
|
iop "iop/proto/gen/iop"
|
|
)
|
|
|
|
// startEventWatcher subscribes to all run and node events, releasing in-flight
|
|
// slots when runs terminate or nodes disconnect. Returns a stop function.
|
|
func (m *modelQueueManager) startEventWatcher(bus *edgeevents.Bus) func() {
|
|
runCh, unsubRun := bus.SubscribeAllRuns(256)
|
|
nodeCh, unsubNode := bus.SubscribeAllNodes(64)
|
|
done := make(chan struct{})
|
|
go func() {
|
|
defer unsubRun()
|
|
defer unsubNode()
|
|
for {
|
|
select {
|
|
case e, ok := <-runCh:
|
|
if !ok {
|
|
return
|
|
}
|
|
if isTerminalRunEvent(e) {
|
|
m.releaseRun(e.GetRunId(), e.GetType())
|
|
}
|
|
case e, ok := <-nodeCh:
|
|
if !ok {
|
|
return
|
|
}
|
|
if e.GetType() == eventpkg.TypeNodeDisconnected {
|
|
m.releaseNode(e.GetNodeId(), "disconnected")
|
|
}
|
|
case <-done:
|
|
return
|
|
}
|
|
}
|
|
}()
|
|
return func() { close(done) }
|
|
}
|
|
|
|
func isTerminalRunEvent(e *iop.RunEvent) bool {
|
|
t := e.GetType()
|
|
return t == "complete" || t == "error" || t == "cancelled"
|
|
}
|
|
|
|
// releaseRun releases the in-flight slot for a terminated run and dispatches
|
|
// the next queued item if one is waiting.
|
|
func (m *modelQueueManager) releaseRun(runID, reason string) {
|
|
m.mu.Lock()
|
|
rec, ok := m.inflightByRun[runID]
|
|
if !ok {
|
|
m.mu.Unlock()
|
|
return
|
|
}
|
|
delete(m.inflightByRun, runID)
|
|
m.releaseSlotLocked(rec.groupKey, rec.nodeID, rec.providerID, rec.long)
|
|
m.mu.Unlock()
|
|
}
|
|
|
|
// releaseNode resets all in-flight slots on a disconnected node, removes it
|
|
// from all queued items' candidate lists, and tries to re-dispatch to any
|
|
// remaining live candidates.
|
|
func (m *modelQueueManager) releaseNode(nodeID, reason string) {
|
|
m.mu.Lock()
|
|
defer m.mu.Unlock()
|
|
|
|
for runID, rec := range m.inflightByRun {
|
|
if rec.nodeID == nodeID {
|
|
delete(m.inflightByRun, runID)
|
|
}
|
|
}
|
|
|
|
for _, group := range m.groups {
|
|
// Remove the disconnected node from all queued items so a future
|
|
// tryDispatch cannot pick it as a stale available candidate.
|
|
for _, item := range group.queue {
|
|
filtered := make([]candidateNode, 0, len(item.candidates))
|
|
for _, c := range item.candidates {
|
|
if c.entry.NodeID != nodeID {
|
|
filtered = append(filtered, c)
|
|
}
|
|
}
|
|
item.candidates = filtered
|
|
}
|
|
|
|
// Clear all slot keys that start with this nodeID.
|
|
// For provider-pool: "nodeID:providerID". For legacy: "nodeID".
|
|
// Long-context in-flight counters are keyed identically, so clear them
|
|
// alongside the normal counters.
|
|
for slot := range group.inflight {
|
|
if colonIdx := findLastColon(slot); colonIdx > 0 {
|
|
candidateNodeID := slot[:colonIdx]
|
|
// Only clear if the node part matches.
|
|
if candidateNodeID == nodeID {
|
|
delete(group.inflight, slot)
|
|
delete(group.longInflight, slot)
|
|
continue
|
|
}
|
|
} else {
|
|
// Legacy slot: exact nodeID match.
|
|
if slot == nodeID {
|
|
delete(group.inflight, slot)
|
|
delete(group.longInflight, slot)
|
|
}
|
|
}
|
|
}
|
|
|
|
// tryDispatchLocked will handle remaining queued items, but we need
|
|
// to call it if we cleared any inflight.
|
|
// Since we already deleted entries, tryDispatchLocked will find available slots.
|
|
// We need to re-check: if any slots were cleared and there are queued items.
|
|
if len(group.queue) > 0 {
|
|
m.tryDispatchLocked(group)
|
|
}
|
|
}
|
|
}
|
|
|
|
// releaseSlot decrements the in-flight count and tries to dispatch the next
|
|
// queued item. Used when admit-path I/O fails after the slot was reserved.
|
|
// For provider-pool dispatches (nodeID:providerID slot), pass providerID.
|
|
// For legacy dispatches, pass empty providerID so nodeID is used directly.
|
|
func (m *modelQueueManager) releaseSlot(groupKey, nodeID string, providerID ...string) {
|
|
var pid string
|
|
if len(providerID) > 0 {
|
|
pid = providerID[0]
|
|
}
|
|
m.releaseSlotWithLong(groupKey, nodeID, pid, false)
|
|
}
|
|
|
|
// releaseSlotWithLong releases a reserved slot, additionally freeing a long-context
|
|
// slot when longSlot is true. Used by admit-path failure/race handling where the
|
|
// caller knows whether a long slot was reserved.
|
|
func (m *modelQueueManager) releaseSlotWithLong(groupKey, nodeID, providerID string, longSlot bool) {
|
|
m.mu.Lock()
|
|
m.releaseSlotLocked(groupKey, nodeID, providerID, longSlot)
|
|
m.mu.Unlock()
|
|
}
|
|
|
|
func (m *modelQueueManager) releaseSlotLocked(groupKey, nodeID string, providerID string, longSlot bool) {
|
|
group, ok := m.groups[groupKey]
|
|
if !ok {
|
|
return
|
|
}
|
|
// Use provider-aware slot key for provider-pool dispatches.
|
|
slot := nodeID
|
|
if providerID != "" {
|
|
slot = nodeID + ":" + providerID
|
|
}
|
|
if group.inflight[slot] > 0 {
|
|
group.inflight[slot]--
|
|
}
|
|
if longSlot && group.longInflight[slot] > 0 {
|
|
group.longInflight[slot]--
|
|
}
|
|
m.tryDispatchLocked(group)
|
|
}
|