mirror of
https://github.com/hanzoai/tasks.git
synced 2026-10-04 02:04:03 +00:00
Lina JodoinandGitHub
a646b2d5ef
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.
55 lines
1.7 KiB
Go
55 lines
1.7 KiB
Go
package scheduler
|
|
|
|
import (
|
|
"slices"
|
|
"unicode/utf8"
|
|
|
|
"go.temporal.io/server/chasm"
|
|
"go.temporal.io/server/chasm/lib/scheduler/gen/schedulerpb/v1"
|
|
"google.golang.org/protobuf/types/known/timestamppb"
|
|
)
|
|
|
|
// EventLog is a CHASM component that keeps a bounded, human-readable history
|
|
// of state changes for its parent component. Entries are not used for any
|
|
// scheduler computation.
|
|
type EventLog struct {
|
|
chasm.UnimplementedComponent
|
|
|
|
*schedulerpb.EventLog
|
|
}
|
|
|
|
// NewEventLog returns an initialized EventLog component, intended to be parented
|
|
// under any component that wants to record events.
|
|
func NewEventLog(ctx chasm.MutableContext) *EventLog {
|
|
return &EventLog{
|
|
EventLog: &schedulerpb.EventLog{},
|
|
}
|
|
}
|
|
|
|
func (e *EventLog) LifecycleState(ctx chasm.Context) chasm.LifecycleState {
|
|
return chasm.LifecycleStateRunning
|
|
}
|
|
|
|
// LogEvent appends an event with the given message. Messages longer than
|
|
// maxMessageLen bytes are truncated at a UTF-8 rune boundary; once the log
|
|
// has more than maxEntries entries, the earliest entries are dropped.
|
|
func (e *EventLog) LogEvent(ctx chasm.MutableContext, msg string, maxEntries, maxMessageLen int) {
|
|
if len(msg) > maxMessageLen {
|
|
// Back off to the nearest UTF-8 rune boundary so we don't split a
|
|
// multibyte rune.
|
|
truncateAt := maxMessageLen
|
|
for truncateAt > 0 && !utf8.RuneStart(msg[truncateAt]) {
|
|
truncateAt--
|
|
}
|
|
msg = msg[:truncateAt]
|
|
}
|
|
e.Events = append(e.Events, &schedulerpb.Event{
|
|
Time: timestamppb.New(ctx.Now(e)),
|
|
Message: msg,
|
|
})
|
|
if keepFrom := len(e.Events) - maxEntries; keepFrom > 0 {
|
|
// Clone so the dropped entries don't stay reachable via the backing array.
|
|
e.Events = slices.Clone(e.Events[keepFrom:])
|
|
}
|
|
}
|