Files
tasks/chasm/lib/scheduler/generator.go
Lina JodoinandGitHub a646b2d5ef [Scheduler] Introduce EventLog for internal history (#10258)
I'm sending the review for the mechanism (and an example callsite), and
I'll send out a review that makes use of it throughout the scheduler
code afterwards.

## What changed?
- CHASM scheduler components each parent a new EventLog component, which
we'll use to record decision logic details that are normally too spammy
for logs.
- No plan to expose this through any APIs just yet; it isn't meant to be
customer-facing (though, I am thinking about how we can make it easy for
dev success to get at)

## Why?
- V1 debugging was often done through use of the workflow's event
history; without an analogue for V2, it's difficult to get a clear
picture of how the logic arrived at a certain output state.
- Info logging in the same capacity would likely be too spammy for our
observability system.
- It's difficult to trace a complete picture for an individual schedule
*over time* using just Loki. Schedules are often long-lived, with state
transitions taking place in short blips over periods of days or weeks.
Being able to get a complete list on demand for a relevant scheduler
will greatly improve debugging time.

## How did you test it?
- [ ] built
- [ ] run locally and tested manually
- [ ] covered by existing tests
- [x] added new unit test(s)
- [ ] added new functional test(s)

## Potential risks
- This could increase scheduler state size, however:
- V2 already measures notably lighter on storage than V1 (not to
mention, no event history data whatsoever)
- Log is bounded to max 30 messages per component (usually scheduler,
generator, invoker) of max 1000 characters each (e.g., ~100KB), as a
safety mechanism
- Most messages logged will be small and informational, not large state
dumps (e.g., "scheduler unpaused", "generator scheduled for ...").
- Backfillers each get their own log, which shares a Backfiller's
lifecycle.
2026-05-27 15:21:52 -07:00

93 lines
2.7 KiB
Go

package scheduler
import (
"time"
"go.temporal.io/server/chasm"
"go.temporal.io/server/chasm/lib/scheduler/gen/schedulerpb/v1"
"go.temporal.io/server/service/worker/scheduler"
"google.golang.org/protobuf/types/known/timestamppb"
)
// The Generator component is responsible for buffering actions according
// to the schedule's specification. Manually requested actions (from an immediate
// request or backfill) are separately handled in the Backfiller component.
type Generator struct {
chasm.UnimplementedComponent
*schedulerpb.GeneratorState
Scheduler chasm.ParentPtr[*Scheduler]
EventLog chasm.Field[*EventLog]
}
// NewGenerator returns an initialized Generator component, which should
// be parented under a Scheduler root node.
func NewGenerator(ctx chasm.MutableContext) *Generator {
generator := newGeneratorWithState(ctx, &schedulerpb.GeneratorState{
LastProcessedTime: nil,
})
// Kick off initial generator run as an immediate task.
generator.Generate(ctx)
return generator
}
func newGeneratorWithState(ctx chasm.MutableContext, state *schedulerpb.GeneratorState) *Generator {
generator := &Generator{
GeneratorState: state,
EventLog: chasm.NewComponentField(ctx, NewEventLog(ctx)),
}
return generator
}
// Generate immediately kicks off a new GeneratorTask. Used after updating the
// schedule specification.
func (g *Generator) Generate(ctx chasm.MutableContext) {
g.scheduleTask(ctx, chasm.TaskScheduledTimeImmediate)
}
// scheduleTask schedules a GeneratorTask at the given time.
func (g *Generator) scheduleTask(ctx chasm.MutableContext, scheduledTime time.Time) {
ctx.AddTask(g, chasm.TaskAttributes{
ScheduledTime: scheduledTime,
}, &schedulerpb.GeneratorTask{})
}
func (g *Generator) LifecycleState(ctx chasm.Context) chasm.LifecycleState {
return chasm.LifecycleStateRunning
}
// UpdateFutureActionTimes computes and stores the next scheduled action times.
func (g *Generator) UpdateFutureActionTimes(
ctx chasm.Context,
specBuilder *scheduler.SpecBuilder,
) {
sched := g.Scheduler.Get(ctx)
spec, err := sched.getCompiledSpec(specBuilder)
if err != nil {
return
}
count := recentActionCount
if sched.Schedule.State.LimitedActions {
count = min(int(sched.Schedule.State.RemainingActions), recentActionCount)
}
futureTimes := make([]*timestamppb.Timestamp, 0, count)
// Start from max(now, updateTime) to ensure we skip times before the last update.
t := ctx.Now(g)
if updateTime := sched.Info.GetUpdateTime().AsTime(); updateTime.After(t) {
t = updateTime
}
for len(futureTimes) < count {
t = spec.GetNextTime(sched.jitterSeed(), t).Next
if t.IsZero() {
break
}
futureTimes = append(futureTimes, timestamppb.New(t))
}
g.FutureActionTimes = futureTimes
}