lostluck commented on code in PR #33763:
URL: https://github.com/apache/beam/pull/33763#discussion_r1940527641
##########
sdks/go/pkg/beam/runners/prism/internal/engine/strategy.go:
##########
@@ -23,16 +23,553 @@ import (
"github.com/apache/beam/sdks/v2/go/pkg/beam/core/typex"
)
+// This file is intended to consolidate handling of WindowingStrategy and
trigger
+// logic independantly from the bulk of the ElementManager.
+//
+// Triggers by their nature only apply to Aggregation transforms, in
particular:
+// GroupByKeys and CoGroupByKeys are aggregations.
+// Triggers also affect downstream side inputs. That is, a side input consumer
+// is vaccuously an aggregation.
+//
+// Triggers are PerKey+PerWindow, and they may or may not use state per
trigger.
+//
+// The unique state per trigger is the trickiest bit to handle. In principle
+// it could just be handled by the existing state system, which might be
+// sufficient, but it's not prepared for unique tagging per trigger itself.
+// It would also add additional overhead since state is kept as the serialized
+// bytes, instead of in a manipulatable form.
+//
+// Instead, each key+window state cell contains a trigger specific state map,
+// handled via pointer equality from the trigger itself.
+
// WinStrat configures the windowing strategy for the stage, based on the
// stage's input PCollection.
type WinStrat struct {
AllowedLateness time.Duration // Used to extend duration
+
+ Trigger Trigger // Evaluated during execution.
+}
+
+// IsTriggerReady updates the trigger state with the given input, and returns
+// if the trigger is ready to fire.
+func (ws WinStrat) IsTriggerReady(input triggerInput, state *StateData) bool {
+ ws.Trigger.onElement(input, state)
+
+ if ws.Trigger.shouldFire(state) {
+ ws.Trigger.onFire(state)
+ return true
+ }
+ return false
}
+// EarliestCompletion marks when we can close a window.
func (ws WinStrat) EarliestCompletion(w typex.Window) mtime.Time {
return w.MaxTimestamp().Add(ws.AllowedLateness)
}
+func (ws WinStrat) IsNeverTrigger() bool {
+ _, ok := ws.Trigger.(*TriggerNever)
+ return ok
+}
+
func (ws WinStrat) String() string {
- return fmt.Sprintf("WinStrat[AllowedLateness:%v]", ws.AllowedLateness)
+ return fmt.Sprintf("WinStrat[AllowedLateness:%v Trigger:%v]",
ws.AllowedLateness, ws.Trigger)
+}
+
+// triggerInput represents a Key + window + stage's trigger conditions.
+type triggerInput struct {
+ newElementCount int // The number of new elements since the last
check.
+ endOfWindowReached bool // Whether or not the end of the window has
been reached.
+}
+
+// Trigger represents a trigger for a windowing strategy. A trigger
determines when
+// to fire a window based on the arrival of elements and the passage of time.
+//
+// See https://s.apache.org/beam-triggers for a more detailed look at triggers.
+type Trigger interface {
+ reset(state *StateData)
+
+ // onElement updates the trigger state based on the provided input.
This may
+ // transition triggers into a fireable state, but will never make them
"finished".
+ onElement(input triggerInput, state *StateData)
+ // shouldFire returns whether the trigger is able to fire or not.
+ shouldFire(state *StateData) bool
+ // onFire commits that the trigger has fired, so triggers may
transition to
+ // a finished state.
+ onFire(state *StateData)
+
+ // TODO handle https://github.com/apache/beam/issues/31438 merging
triggers and state for merging windows (sessions, but also custom merging
windows)
+}
+
+// triggerState retains additional state for a given trigger execution.
+// Each trigger is responsible for maintaining it's own state as needed.
+type triggerState struct {
+ // finished indicates if the trigger has already fired or not.
+ finished bool
+ // extra is where additional data can be stored.
+ extra any
+}
+
+func (ts triggerState) String() string {
+ return fmt.Sprintf("triggerState[finished: %v; state: %v]",
ts.finished, ts.extra)
+}
+
+// nullTrigger is a 0 size object that exists to be embedded in triggers that
+// perform no action on trigger method calls. Triggers with this embedded will
+// gain an implementation of the trigger methods that do nothing, and behavior
+// must be overridden by the trigger for correct evaluation.
+type nullTrigger struct{}
+
+func (nullTrigger) onElement(triggerInput, *StateData) {}
+func (nullTrigger) onFire(*StateData) {}
+func (nullTrigger) reset(*StateData) {}
+
+// TriggerNever is never ready.
+// There will only be an ON_TIME output and a final output at window
expiration.
+type TriggerNever struct{ nullTrigger }
+
+func (*TriggerNever) shouldFire(*StateData) bool {
+ return false
+}
+
+func (t *TriggerNever) reset(state *StateData) {}
+
+func (t *TriggerNever) String() string {
+ return "Never"
+}
+
+// TriggerAlways is always ready.
+// There will be an output for every element, and a final output at window
expiration.
+// Equivalent to TriggerRepeatedly {TriggerElementCount{1}}
+type TriggerAlways struct{ nullTrigger }
+
+func (*TriggerAlways) shouldFire(*StateData) bool {
+ return true
+}
+
+func subTriggersOnElement(t Trigger, input triggerInput, state *StateData,
subTriggers []Trigger) {
+ ts := state.getTriggerState(t)
+ if ts.finished {
+ return
+ }
+
+ for _, sub := range subTriggers {
+ sub.onElement(input, state)
+ }
+}
+
+func subTriggersReset(t Trigger, state *StateData, subTriggers []Trigger) {
+ for _, sub := range subTriggers {
+ sub.reset(state)
+ }
+ delete(state.Trigger, t)
+}
+
+func triggerClearAndFinish(t Trigger, state *StateData) {
+ t.reset(state)
+ ts := state.getTriggerState(t)
+ ts.finished = true
+ state.setTriggerState(t, ts)
+}
+
+func (t *TriggerAlways) String() string {
+ return "Always"
+}
+
+// TriggerAfterAll is ready when all subTriggers are ready.
+// There will be an output when all subTriggers are ready.
+// Logically, an "AND" trigger.
+type TriggerAfterAll struct {
+ SubTriggers []Trigger
+}
+
+func (t *TriggerAfterAll) onElement(input triggerInput, state *StateData) {
+ subTriggersOnElement(t, input, state, t.SubTriggers)
+}
+
+func (t *TriggerAfterAll) shouldFire(state *StateData) bool {
+ ts := state.getTriggerState(t)
+ if ts.finished {
+ return false
+ }
+ shouldFire := true
+ for _, sub := range t.SubTriggers {
+ shouldFire = shouldFire && sub.shouldFire(state)
+ }
+ return shouldFire
+}
+
+func (t *TriggerAfterAll) onFire(state *StateData) {
+ unfinished := false
+ for _, sub := range t.SubTriggers {
+ if sub.shouldFire(state) {
+ sub.onFire(state)
+ }
+ if !state.getTriggerState(sub).finished {
+ unfinished = true
+ }
+ }
+ if unfinished {
+ return
+ }
+ triggerClearAndFinish(t, state)
+}
+
+func (t *TriggerAfterAll) reset(state *StateData) {
+ subTriggersReset(t, state, t.SubTriggers)
+}
+
+func (t *TriggerAfterAll) String() string {
+ return fmt.Sprintf("AfterAll[%v]", t.SubTriggers)
+}
+
+// TriggerAfterAny is ready the first time any of the subTriggers are ready.
+// Logically, an "OR" trigger.
+type TriggerAfterAny struct {
+ SubTriggers []Trigger
+}
+
+func (t *TriggerAfterAny) onElement(input triggerInput, state *StateData) {
+ subTriggersOnElement(t, input, state, t.SubTriggers)
+}
+
+func (t *TriggerAfterAny) shouldFire(state *StateData) bool {
+ ts := state.getTriggerState(t)
+ if ts.finished {
+ return false
+ }
+ for _, sub := range t.SubTriggers {
+ if sub.shouldFire(state) {
+ return true
+ }
+ }
+ return false
+}
+
+func (t *TriggerAfterAny) onFire(state *StateData) {
+ if !t.shouldFire(state) {
+ return
+ }
+ triggerClearAndFinish(t, state)
+}
+
+func (t *TriggerAfterAny) reset(state *StateData) {
+ subTriggersReset(t, state, t.SubTriggers)
+}
+
+func (t *TriggerAfterAny) String() string {
+ return fmt.Sprintf("AfterAny[%v]", t.SubTriggers)
+}
+
+// TriggerAfterEach processes each trigger before executing the next.
+// Starting with the first subtrigger, ready when the _current_ subtrigger
+// is ready. After output, advances the current trigger by one.
+type TriggerAfterEach struct {
+ SubTriggers []Trigger
+}
+
+func (t *TriggerAfterEach) onElement(input triggerInput, state *StateData) {
+ ts := state.getTriggerState(t)
+ if ts.finished {
+ return
+ }
+ // Only update the first unfinished sub trigger.
+ for _, sub := range t.SubTriggers {
+ if state.getTriggerState(sub).finished {
+ continue
+ }
+ sub.onElement(input, state)
+ return
+ }
+}
+
+func (t *TriggerAfterEach) shouldFire(state *StateData) bool {
+ ts := state.getTriggerState(t)
+ if ts.finished {
+ return false
+ }
+ for _, sub := range t.SubTriggers {
+ if state.getTriggerState(sub).finished {
+ continue
+ }
+ return sub.shouldFire(state)
+ }
+ return false
+}
+
+func (t *TriggerAfterEach) onFire(state *StateData) {
+ if !t.shouldFire(state) {
Review Comment:
It's largely that at this point I'd rather have an implementation that
works, than be efficient.
If efficiency is required, those can be much smaller targetted PRs, as
guided by a profiler.
##########
sdks/go/pkg/beam/runners/prism/internal/engine/strategy.go:
##########
@@ -23,16 +23,553 @@ import (
"github.com/apache/beam/sdks/v2/go/pkg/beam/core/typex"
)
+// This file is intended to consolidate handling of WindowingStrategy and
trigger
+// logic independantly from the bulk of the ElementManager.
+//
+// Triggers by their nature only apply to Aggregation transforms, in
particular:
+// GroupByKeys and CoGroupByKeys are aggregations.
+// Triggers also affect downstream side inputs. That is, a side input consumer
+// is vaccuously an aggregation.
+//
+// Triggers are PerKey+PerWindow, and they may or may not use state per
trigger.
+//
+// The unique state per trigger is the trickiest bit to handle. In principle
+// it could just be handled by the existing state system, which might be
+// sufficient, but it's not prepared for unique tagging per trigger itself.
+// It would also add additional overhead since state is kept as the serialized
+// bytes, instead of in a manipulatable form.
+//
+// Instead, each key+window state cell contains a trigger specific state map,
+// handled via pointer equality from the trigger itself.
+
// WinStrat configures the windowing strategy for the stage, based on the
// stage's input PCollection.
type WinStrat struct {
AllowedLateness time.Duration // Used to extend duration
+
+ Trigger Trigger // Evaluated during execution.
+}
+
+// IsTriggerReady updates the trigger state with the given input, and returns
+// if the trigger is ready to fire.
Review Comment:
I think I've always understood "ready to fire" here means "Caller, you must
create a bundle now".
Again, it's an efficiency thing more than anything. The subsequent PR
creates the triggered bundle immeadiately, which is a valid approach to how the
triggers fire.
It could instead finish processing all the elements received, and then fire
a larger consolidated bundle for the firing.
But that leads to performance optimization which I sadly have no time for.
##########
sdks/go/pkg/beam/runners/prism/internal/engine/strategy.go:
##########
@@ -23,16 +23,496 @@ import (
"github.com/apache/beam/sdks/v2/go/pkg/beam/core/typex"
)
+// This file is intended to consolidate handling of WindowingStrategy and
trigger
+// logic independantly from the bulk of the ElementManager.
+//
+// Triggers by their nature only apply to Aggregation transforms, in
particular:
+// GroupByKeys and CoGroupByKeys are aggregations.
+// Triggers also affect downstream side inputs. That is, a side input consumer
+// is vaccuously an aggregation.
+//
+// Triggers are PerKey+PerWindow, and they may or may not use state per
trigger.
+//
+// The unique state per trigger is the trickiest bit to handle. In principle
+// it could just be handled by the existing state system, which might be
+// sufficient, but it's not prepared for unique tagging per trigger itself.
+// It would also add additional overhead since state is kept as the serialized
+// bytes, instead of in a manipulatable form.
+//
+// Instead, each key+window state cell contains a trigger specific state map,
+// handled via pointer equality from the trigger itself.
+
// WinStrat configures the windowing strategy for the stage, based on the
// stage's input PCollection.
type WinStrat struct {
AllowedLateness time.Duration // Used to extend duration
+
+ Trigger Trigger // Evaluated during execution.
}
+// IsTriggerReady updates and returns wether the trigger is able to fire or
not.
+func (ws WinStrat) IsTriggerReady(input triggerInput, state *StateData) bool {
+ ws.Trigger.onElement(input, state)
+
+ if ws.Trigger.shouldFire(state) {
+ ws.Trigger.onFire(state)
+ return true
+ }
+ return false
+}
+
+// EarliestCompletion marks when we can close a window.
func (ws WinStrat) EarliestCompletion(w typex.Window) mtime.Time {
return w.MaxTimestamp().Add(ws.AllowedLateness)
}
func (ws WinStrat) String() string {
return fmt.Sprintf("WinStrat[AllowedLateness:%v]", ws.AllowedLateness)
}
+
+// triggerInput represents a Key + window + stage's trigger conditions.
+type triggerInput struct {
+ newElementCount int // The number of new elements since the last
check.
+ endOfWindowReached bool // Whether or not the end of the window has
been reached.
+}
+
+// Trigger represents a trigger for a windowing strategy. A trigger
determines when
+// to fire a window based on the arrival of elements and the passage of time.
+//
+// See https://s.apache.org/beam-triggers for a more detailed look at triggers.
+type Trigger interface {
+ reset(state *StateData)
+
+ // onElement updates the trigger state based on the provided input.
This may
+ // transition triggers into a fireable state, but will never make them
"finished".
+ onElement(input triggerInput, state *StateData)
+ // shouldFire returns whether the trigger is able to fire or not.
+ shouldFire(state *StateData) bool
+ // onFire commits that the trigger has fired, so triggers may
transition to
+ // a finished state.
+ onFire(state *StateData)
+
+ // TODO merging triggers and state for merging windows
+}
+
+// triggerState retains additional state for a given trigger execution.
+// Each trigger is responsible for maintaining it's own state as needed.
+type triggerState struct {
+ // finished indicates if the trigger has already fired or not.
+ finished bool
+ // extra is where additional data can be stored.
+ extra any
+}
+
+// nullTrigger is a 0 size object that exists to be embedded in triggers that
+// perform no action on trigger method calls. Triggers with this embedded will
+// gain an implementation of the trigger methods that do nothing, and behavior
+// must be overridden by the trigger for correct evaluation.
+type nullTrigger struct{}
+
+func (nullTrigger) onElement(triggerInput, *StateData) {}
+func (nullTrigger) onFire(*StateData) {}
+func (nullTrigger) reset(*StateData) {}
+
+// TriggerNever is never ready.
+// There will only be an ON_TIME output and a final output at window
expiration.
+type TriggerNever struct{ nullTrigger }
+
+func (*TriggerNever) shouldFire(*StateData) bool {
+ return false
+}
+
+// TriggerAlways is always ready.
+// There will be an output for every element, and a final output at window
expiration.
+// Equivalent to TriggerRepeatedly {TriggerElementCount{1}}
+type TriggerAlways struct{ nullTrigger }
+
+func (*TriggerAlways) shouldFire(*StateData) bool {
+ return true
+}
+
+func subTriggersOnElement(t Trigger, input triggerInput, state *StateData,
subTriggers []Trigger) {
+ ts := state.getTriggerState(t)
+ if ts.finished {
+ return
+ }
+
+ for _, sub := range subTriggers {
+ sub.onElement(input, state)
+ }
+}
+
+func subTriggersReset(t Trigger, state *StateData, subTriggers []Trigger) {
+ for _, sub := range subTriggers {
+ sub.reset(state)
+ }
+ delete(state.Trigger, t)
+}
+
+func triggerClearAndFinish(t Trigger, state *StateData) {
+ t.reset(state)
+ ts := state.getTriggerState(t)
+ ts.finished = true
+ state.setTriggerState(t, ts)
+}
+
+// TriggerAfterAll is ready when all subTriggers are ready.
+// There will be an output when all subTriggers are ready.
+// Logically, an "AND" trigger.
+type TriggerAfterAll struct {
+ SubTriggers []Trigger
+}
+
+func (t *TriggerAfterAll) onElement(input triggerInput, state *StateData) {
+ subTriggersOnElement(t, input, state, t.SubTriggers)
+}
+
+func (t *TriggerAfterAll) shouldFire(state *StateData) bool {
+ ts := state.getTriggerState(t)
+ if ts.finished {
+ return false
+ }
+ shouldFire := true
+ for _, sub := range t.SubTriggers {
+ shouldFire = shouldFire && sub.shouldFire(state)
+ }
+ return shouldFire
+}
+
+func (t *TriggerAfterAll) onFire(state *StateData) {
+ unfinished := false
+ for _, sub := range t.SubTriggers {
+ if sub.shouldFire(state) {
+ sub.onFire(state)
+ }
+ if !state.getTriggerState(sub).finished {
+ unfinished = true
+ }
+ }
+ if unfinished {
+ return
+ }
+ triggerClearAndFinish(t, state)
+}
+
+func (t *TriggerAfterAll) reset(state *StateData) {
+ subTriggersReset(t, state, t.SubTriggers)
+}
+
+// TriggerAfterAny is ready the first time any of the subTriggers are ready.
+// Logically, an "OR" trigger.
+type TriggerAfterAny struct {
+ SubTriggers []Trigger
+}
+
+func (t *TriggerAfterAny) onElement(input triggerInput, state *StateData) {
+ subTriggersOnElement(t, input, state, t.SubTriggers)
+}
+
+func (t *TriggerAfterAny) shouldFire(state *StateData) bool {
+ ts := state.getTriggerState(t)
+ if ts.finished {
+ return false
+ }
+ for _, sub := range t.SubTriggers {
+ if sub.shouldFire(state) {
+ return true
+ }
+ }
+ return false
+}
+
+func (t *TriggerAfterAny) onFire(state *StateData) {
+ if !t.shouldFire(state) {
+ return
+ }
+ triggerClearAndFinish(t, state)
+}
+
+func (t *TriggerAfterAny) reset(state *StateData) {
+ subTriggersReset(t, state, t.SubTriggers)
+}
+
+// TriggerAfterEach processes each trigger before executing the next.
+// Starting with the first subtrigger, ready when the _current_ subtrigger
+// is ready. After output, advances the current trigger by one.
+type TriggerAfterEach struct {
+ SubTriggers []Trigger
+}
+
+func (t *TriggerAfterEach) onElement(input triggerInput, state *StateData) {
+ ts := state.getTriggerState(t)
+ if ts.finished {
+ return
+ }
+ // Only update the first unfinished sub trigger.
+ for _, sub := range t.SubTriggers {
+ if state.getTriggerState(sub).finished {
+ continue
+ }
+ sub.onElement(input, state)
+ return
+ }
+}
+
+func (t *TriggerAfterEach) shouldFire(state *StateData) bool {
+ ts := state.getTriggerState(t)
+ if ts.finished {
+ return false
+ }
+ for _, sub := range t.SubTriggers {
+ if state.getTriggerState(sub).finished {
+ continue
+ }
+ return sub.shouldFire(state)
+ }
+ return false
+}
+
+func (t *TriggerAfterEach) onFire(state *StateData) {
+ if !t.shouldFire(state) {
+ return
+ }
+ for _, sub := range t.SubTriggers {
+ if state.getTriggerState(sub).finished {
+ continue
+ }
+ sub.onFire(state)
+ if !state.getTriggerState(sub).finished {
+ return
+ }
+ }
+ triggerClearAndFinish(t, state)
+}
+
+func (t *TriggerAfterEach) reset(state *StateData) {
+ subTriggersReset(t, state, t.SubTriggers)
+}
+
+// TriggerElementCount triggers when there have been at least the required
number
+// of elements have arrived.
+//
+// TriggerElementCount stores the current element count in it's extra state
field.
+type TriggerElementCount struct {
+ ElementCount int
+}
+
+func (t *TriggerElementCount) onElement(input triggerInput, state *StateData) {
+ ts := state.getTriggerState(t)
+ if ts.finished {
+ return
+ }
+
+ if ts.extra == nil {
+ ts.extra = int(0)
+ }
+ count := ts.extra.(int) + input.newElementCount
+ ts.extra = count
+ state.setTriggerState(t, ts)
+}
+
+func (t *TriggerElementCount) shouldFire(state *StateData) bool {
+ ts := state.getTriggerState(t)
+ if ts.finished {
+ return false
+ }
+ if ts.extra == nil {
+ return false
+ }
+ return ts.extra.(int) >= t.ElementCount
Review Comment:
I agree that we should either have ElementCount(0) behave as either the
Never Trigger, or the Always Trigger.
Always basically means in the undefined state case, we return true instead
of false.
##########
sdks/go/pkg/beam/runners/prism/internal/engine/strategy.go:
##########
@@ -23,16 +23,553 @@ import (
"github.com/apache/beam/sdks/v2/go/pkg/beam/core/typex"
)
+// This file is intended to consolidate handling of WindowingStrategy and
trigger
+// logic independantly from the bulk of the ElementManager.
+//
+// Triggers by their nature only apply to Aggregation transforms, in
particular:
+// GroupByKeys and CoGroupByKeys are aggregations.
+// Triggers also affect downstream side inputs. That is, a side input consumer
+// is vaccuously an aggregation.
+//
+// Triggers are PerKey+PerWindow, and they may or may not use state per
trigger.
+//
+// The unique state per trigger is the trickiest bit to handle. In principle
+// it could just be handled by the existing state system, which might be
+// sufficient, but it's not prepared for unique tagging per trigger itself.
+// It would also add additional overhead since state is kept as the serialized
+// bytes, instead of in a manipulatable form.
+//
+// Instead, each key+window state cell contains a trigger specific state map,
+// handled via pointer equality from the trigger itself.
+
// WinStrat configures the windowing strategy for the stage, based on the
// stage's input PCollection.
type WinStrat struct {
AllowedLateness time.Duration // Used to extend duration
+
+ Trigger Trigger // Evaluated during execution.
+}
+
+// IsTriggerReady updates the trigger state with the given input, and returns
+// if the trigger is ready to fire.
+func (ws WinStrat) IsTriggerReady(input triggerInput, state *StateData) bool {
+ ws.Trigger.onElement(input, state)
+
+ if ws.Trigger.shouldFire(state) {
+ ws.Trigger.onFire(state)
+ return true
+ }
+ return false
}
+// EarliestCompletion marks when we can close a window.
func (ws WinStrat) EarliestCompletion(w typex.Window) mtime.Time {
return w.MaxTimestamp().Add(ws.AllowedLateness)
}
+func (ws WinStrat) IsNeverTrigger() bool {
+ _, ok := ws.Trigger.(*TriggerNever)
+ return ok
+}
+
func (ws WinStrat) String() string {
- return fmt.Sprintf("WinStrat[AllowedLateness:%v]", ws.AllowedLateness)
+ return fmt.Sprintf("WinStrat[AllowedLateness:%v Trigger:%v]",
ws.AllowedLateness, ws.Trigger)
+}
+
+// triggerInput represents a Key + window + stage's trigger conditions.
+type triggerInput struct {
+ newElementCount int // The number of new elements since the last
check.
+ endOfWindowReached bool // Whether or not the end of the window has
been reached.
+}
+
+// Trigger represents a trigger for a windowing strategy. A trigger
determines when
+// to fire a window based on the arrival of elements and the passage of time.
+//
+// See https://s.apache.org/beam-triggers for a more detailed look at triggers.
+type Trigger interface {
+ reset(state *StateData)
+
+ // onElement updates the trigger state based on the provided input.
This may
+ // transition triggers into a fireable state, but will never make them
"finished".
+ onElement(input triggerInput, state *StateData)
+ // shouldFire returns whether the trigger is able to fire or not.
+ shouldFire(state *StateData) bool
+ // onFire commits that the trigger has fired, so triggers may
transition to
+ // a finished state.
+ onFire(state *StateData)
+
+ // TODO handle https://github.com/apache/beam/issues/31438 merging
triggers and state for merging windows (sessions, but also custom merging
windows)
+}
+
+// triggerState retains additional state for a given trigger execution.
+// Each trigger is responsible for maintaining it's own state as needed.
+type triggerState struct {
+ // finished indicates if the trigger has already fired or not.
+ finished bool
+ // extra is where additional data can be stored.
+ extra any
+}
+
+func (ts triggerState) String() string {
+ return fmt.Sprintf("triggerState[finished: %v; state: %v]",
ts.finished, ts.extra)
+}
+
+// nullTrigger is a 0 size object that exists to be embedded in triggers that
+// perform no action on trigger method calls. Triggers with this embedded will
+// gain an implementation of the trigger methods that do nothing, and behavior
+// must be overridden by the trigger for correct evaluation.
+type nullTrigger struct{}
+
+func (nullTrigger) onElement(triggerInput, *StateData) {}
+func (nullTrigger) onFire(*StateData) {}
+func (nullTrigger) reset(*StateData) {}
+
+// TriggerNever is never ready.
+// There will only be an ON_TIME output and a final output at window
expiration.
+type TriggerNever struct{ nullTrigger }
+
+func (*TriggerNever) shouldFire(*StateData) bool {
+ return false
+}
+
+func (t *TriggerNever) reset(state *StateData) {}
+
+func (t *TriggerNever) String() string {
+ return "Never"
+}
+
+// TriggerAlways is always ready.
+// There will be an output for every element, and a final output at window
expiration.
+// Equivalent to TriggerRepeatedly {TriggerElementCount{1}}
+type TriggerAlways struct{ nullTrigger }
+
+func (*TriggerAlways) shouldFire(*StateData) bool {
+ return true
+}
+
+func subTriggersOnElement(t Trigger, input triggerInput, state *StateData,
subTriggers []Trigger) {
+ ts := state.getTriggerState(t)
+ if ts.finished {
+ return
+ }
+
+ for _, sub := range subTriggers {
+ sub.onElement(input, state)
+ }
+}
+
+func subTriggersReset(t Trigger, state *StateData, subTriggers []Trigger) {
+ for _, sub := range subTriggers {
+ sub.reset(state)
+ }
+ delete(state.Trigger, t)
+}
+
+func triggerClearAndFinish(t Trigger, state *StateData) {
+ t.reset(state)
+ ts := state.getTriggerState(t)
+ ts.finished = true
+ state.setTriggerState(t, ts)
+}
+
+func (t *TriggerAlways) String() string {
+ return "Always"
+}
+
+// TriggerAfterAll is ready when all subTriggers are ready.
+// There will be an output when all subTriggers are ready.
+// Logically, an "AND" trigger.
+type TriggerAfterAll struct {
+ SubTriggers []Trigger
+}
+
+func (t *TriggerAfterAll) onElement(input triggerInput, state *StateData) {
+ subTriggersOnElement(t, input, state, t.SubTriggers)
+}
+
+func (t *TriggerAfterAll) shouldFire(state *StateData) bool {
+ ts := state.getTriggerState(t)
+ if ts.finished {
+ return false
+ }
+ shouldFire := true
+ for _, sub := range t.SubTriggers {
+ shouldFire = shouldFire && sub.shouldFire(state)
+ }
+ return shouldFire
+}
+
+func (t *TriggerAfterAll) onFire(state *StateData) {
+ unfinished := false
+ for _, sub := range t.SubTriggers {
+ if sub.shouldFire(state) {
+ sub.onFire(state)
+ }
+ if !state.getTriggerState(sub).finished {
+ unfinished = true
+ }
+ }
+ if unfinished {
+ return
+ }
+ triggerClearAndFinish(t, state)
+}
+
+func (t *TriggerAfterAll) reset(state *StateData) {
+ subTriggersReset(t, state, t.SubTriggers)
+}
+
+func (t *TriggerAfterAll) String() string {
+ return fmt.Sprintf("AfterAll[%v]", t.SubTriggers)
+}
+
+// TriggerAfterAny is ready the first time any of the subTriggers are ready.
+// Logically, an "OR" trigger.
+type TriggerAfterAny struct {
+ SubTriggers []Trigger
+}
+
+func (t *TriggerAfterAny) onElement(input triggerInput, state *StateData) {
+ subTriggersOnElement(t, input, state, t.SubTriggers)
+}
+
+func (t *TriggerAfterAny) shouldFire(state *StateData) bool {
+ ts := state.getTriggerState(t)
+ if ts.finished {
+ return false
+ }
+ for _, sub := range t.SubTriggers {
+ if sub.shouldFire(state) {
+ return true
+ }
+ }
+ return false
+}
+
+func (t *TriggerAfterAny) onFire(state *StateData) {
+ if !t.shouldFire(state) {
+ return
+ }
+ triggerClearAndFinish(t, state)
+}
+
+func (t *TriggerAfterAny) reset(state *StateData) {
+ subTriggersReset(t, state, t.SubTriggers)
+}
+
+func (t *TriggerAfterAny) String() string {
+ return fmt.Sprintf("AfterAny[%v]", t.SubTriggers)
+}
+
+// TriggerAfterEach processes each trigger before executing the next.
+// Starting with the first subtrigger, ready when the _current_ subtrigger
+// is ready. After output, advances the current trigger by one.
+type TriggerAfterEach struct {
+ SubTriggers []Trigger
+}
+
+func (t *TriggerAfterEach) onElement(input triggerInput, state *StateData) {
+ ts := state.getTriggerState(t)
+ if ts.finished {
+ return
+ }
+ // Only update the first unfinished sub trigger.
+ for _, sub := range t.SubTriggers {
+ if state.getTriggerState(sub).finished {
+ continue
+ }
+ sub.onElement(input, state)
+ return
+ }
+}
+
+func (t *TriggerAfterEach) shouldFire(state *StateData) bool {
+ ts := state.getTriggerState(t)
+ if ts.finished {
+ return false
+ }
+ for _, sub := range t.SubTriggers {
+ if state.getTriggerState(sub).finished {
+ continue
+ }
+ return sub.shouldFire(state)
+ }
+ return false
+}
+
+func (t *TriggerAfterEach) onFire(state *StateData) {
+ if !t.shouldFire(state) {
+ return
+ }
+ for _, sub := range t.SubTriggers {
+ if state.getTriggerState(sub).finished {
+ continue
+ }
+ sub.onFire(state)
+ if !state.getTriggerState(sub).finished {
+ return
+ }
+ }
+ triggerClearAndFinish(t, state)
+}
+
+func (t *TriggerAfterEach) reset(state *StateData) {
+ subTriggersReset(t, state, t.SubTriggers)
+}
+
+func (t *TriggerAfterEach) String() string {
+ return fmt.Sprintf("AfterEach[%v]", t.SubTriggers)
+}
+
+// TriggerElementCount triggers when there have been at least the required
number
+// of elements have arrived.
+//
+// TriggerElementCount stores the current element count in it's extra state
field.
+type TriggerElementCount struct {
+ ElementCount int
+}
+
+func (t *TriggerElementCount) onElement(input triggerInput, state *StateData) {
+ ts := state.getTriggerState(t)
+ if ts.finished {
+ return
+ }
+
+ if ts.extra == nil {
+ ts.extra = int(0)
+ }
+ count := ts.extra.(int) + input.newElementCount
+ ts.extra = count
+ state.setTriggerState(t, ts)
+}
+
+func (t *TriggerElementCount) shouldFire(state *StateData) bool {
+ ts := state.getTriggerState(t)
+ if ts.finished {
Review Comment:
I prefer having the explicit check, over the implicit magic.
If each trigger were a leaf rather than an arbitrary tree, then I'd have put
it in a single place external to them.
But the Java approach of spawning manipulation machines from within the
contexts is convoluted indirection IMHO. Being explicit that the trigger is
checking it's finished state is easier for everyone to see wether anything is
handled appropriately in a concrete fashion vs through a 3rd layer abstraction.
##########
sdks/go/pkg/beam/runners/prism/internal/engine/strategy.go:
##########
@@ -23,16 +23,552 @@ import (
"github.com/apache/beam/sdks/v2/go/pkg/beam/core/typex"
)
+// This file is intended to consolidate handling of WindowingStrategy and
trigger
+// logic independantly from the bulk of the ElementManager.
+//
+// Triggers by their nature only apply to Aggregation transforms, in
particular:
+// GroupByKeys and CoGroupByKeys are aggregations.
+// Triggers also affect downstream side inputs. That is, a side input consumer
+// is vaccuously an aggregation.
+//
+// Triggers are PerKey+PerWindow, and they may or may not use state per
trigger.
+//
+// The unique state per trigger is the trickiest bit to handle. In principle
+// it could just be handled by the existing state system, which might be
+// sufficient, but it's not prepared for unique tagging per trigger itself.
+// It would also add additional overhead since state is kept as the serialized
+// bytes, instead of in a manipulatable form.
+//
+// Instead, each key+window state cell contains a trigger specific state map,
+// handled via pointer equality from the trigger itself.
+
// WinStrat configures the windowing strategy for the stage, based on the
// stage's input PCollection.
type WinStrat struct {
AllowedLateness time.Duration // Used to extend duration
+
+ Trigger Trigger // Evaluated during execution.
+}
+
+// IsTriggerReady updates the trigger state with the given input, and returns
+// if the trigger is ready to fire.
+func (ws WinStrat) IsTriggerReady(input triggerInput, state *StateData) bool {
+ ws.Trigger.onElement(input, state)
+
+ if ws.Trigger.shouldFire(state) {
+ ws.Trigger.onFire(state)
+ return true
+ }
+ return false
}
+// EarliestCompletion marks when we can close a window.
func (ws WinStrat) EarliestCompletion(w typex.Window) mtime.Time {
return w.MaxTimestamp().Add(ws.AllowedLateness)
}
+func (ws WinStrat) IsNeverTrigger() bool {
+ _, ok := ws.Trigger.(*TriggerNever)
+ return ok
+}
+
func (ws WinStrat) String() string {
- return fmt.Sprintf("WinStrat[AllowedLateness:%v]", ws.AllowedLateness)
+ return fmt.Sprintf("WinStrat[AllowedLateness:%v Trigger:%v]",
ws.AllowedLateness, ws.Trigger)
+}
+
+// triggerInput represents a Key + window + stage's trigger conditions.
+type triggerInput struct {
+ newElementCount int // The number of new elements since the last
check.
+ endOfWindowReached bool // Whether or not the end of the window has
been reached.
+}
+
+// Trigger represents a trigger for a windowing strategy. A trigger
determines when
+// to fire a window based on the arrival of elements and the passage of time.
+//
+// See https://s.apache.org/beam-triggers for a more detailed look at triggers.
+type Trigger interface {
+ reset(state *StateData)
+
+ // onElement updates the trigger state based on the provided input.
This may
+ // transition triggers into a fireable state, but will never make them
"finished".
+ onElement(input triggerInput, state *StateData)
+ // shouldFire returns whether the trigger is able to fire or not.
+ shouldFire(state *StateData) bool
+ // onFire commits that the trigger has fired, so triggers may
transition to
+ // a finished state.
+ onFire(state *StateData)
+
+ // TODO merging triggers and state for merging windows
+}
+
+// triggerState retains additional state for a given trigger execution.
+// Each trigger is responsible for maintaining it's own state as needed.
+type triggerState struct {
+ // finished indicates if the trigger has already fired or not.
+ finished bool
+ // extra is where additional data can be stored.
+ extra any
+}
+
+func (ts triggerState) String() string {
+ return fmt.Sprintf("triggerState[finished: %v; state: %v]",
ts.finished, ts.extra)
+}
+
+// nullTrigger is a 0 size object that exists to be embedded in triggers that
+// perform no action on trigger method calls. Triggers with this embedded will
+// gain an implementation of the trigger methods that do nothing, and behavior
+// must be overridden by the trigger for correct evaluation.
+type nullTrigger struct{}
+
+func (nullTrigger) onElement(triggerInput, *StateData) {}
+func (nullTrigger) onFire(*StateData) {}
+func (nullTrigger) reset(*StateData) {}
+
+// TriggerNever is never ready.
+// There will only be an ON_TIME output and a final output at window
expiration.
+type TriggerNever struct{ nullTrigger }
+
+func (*TriggerNever) shouldFire(*StateData) bool {
+ return false
+}
+
+func (t *TriggerNever) reset(state *StateData) {}
+
+func (t *TriggerNever) String() string {
+ return "Never"
+}
+
+// TriggerAlways is always ready.
+// There will be an output for every element, and a final output at window
expiration.
+// Equivalent to TriggerRepeatedly {TriggerElementCount{1}}
+type TriggerAlways struct{ nullTrigger }
+
+func (*TriggerAlways) shouldFire(*StateData) bool {
+ return true
+}
+
+func subTriggersOnElement(t Trigger, input triggerInput, state *StateData,
subTriggers []Trigger) {
+ ts := state.getTriggerState(t)
+ if ts.finished {
+ return
+ }
+
+ for _, sub := range subTriggers {
+ sub.onElement(input, state)
+ }
+}
+
+func subTriggersReset(t Trigger, state *StateData, subTriggers []Trigger) {
+ for _, sub := range subTriggers {
+ sub.reset(state)
+ }
+ delete(state.Trigger, t)
+}
+
+func triggerClearAndFinish(t Trigger, state *StateData) {
+ t.reset(state)
+ ts := state.getTriggerState(t)
+ ts.finished = true
+ state.setTriggerState(t, ts)
+}
+
+func (t *TriggerAlways) String() string {
+ return "Always"
+}
+
+// TriggerAfterAll is ready when all subTriggers are ready.
+// There will be an output when all subTriggers are ready.
+// Logically, an "AND" trigger.
+type TriggerAfterAll struct {
+ SubTriggers []Trigger
+}
+
+func (t *TriggerAfterAll) onElement(input triggerInput, state *StateData) {
+ subTriggersOnElement(t, input, state, t.SubTriggers)
+}
+
+func (t *TriggerAfterAll) shouldFire(state *StateData) bool {
+ ts := state.getTriggerState(t)
+ if ts.finished {
+ return false
+ }
+ shouldFire := true
+ for _, sub := range t.SubTriggers {
+ shouldFire = shouldFire && sub.shouldFire(state)
+ }
+ return shouldFire
+}
+
+func (t *TriggerAfterAll) onFire(state *StateData) {
+ unfinished := false
+ for _, sub := range t.SubTriggers {
+ if sub.shouldFire(state) {
+ sub.onFire(state)
+ }
+ if !state.getTriggerState(sub).finished {
+ unfinished = true
+ }
+ }
+ if unfinished {
+ return
+ }
+ triggerClearAndFinish(t, state)
+}
+
+func (t *TriggerAfterAll) reset(state *StateData) {
+ subTriggersReset(t, state, t.SubTriggers)
+}
+
+func (t *TriggerAfterAll) String() string {
+ return fmt.Sprintf("AfterAll[%v]", t.SubTriggers)
+}
+
+// TriggerAfterAny is ready the first time any of the subTriggers are ready.
+// Logically, an "OR" trigger.
+type TriggerAfterAny struct {
+ SubTriggers []Trigger
+}
+
+func (t *TriggerAfterAny) onElement(input triggerInput, state *StateData) {
+ subTriggersOnElement(t, input, state, t.SubTriggers)
+}
+
+func (t *TriggerAfterAny) shouldFire(state *StateData) bool {
+ ts := state.getTriggerState(t)
+ if ts.finished {
+ return false
+ }
+ for _, sub := range t.SubTriggers {
+ if sub.shouldFire(state) {
+ return true
+ }
+ }
+ return false
+}
+
+func (t *TriggerAfterAny) onFire(state *StateData) {
+ if !t.shouldFire(state) {
+ return
+ }
+ triggerClearAndFinish(t, state)
+}
+
+func (t *TriggerAfterAny) reset(state *StateData) {
+ subTriggersReset(t, state, t.SubTriggers)
+}
+
+func (t *TriggerAfterAny) String() string {
+ return fmt.Sprintf("AfterAny[%v]", t.SubTriggers)
+}
+
+// TriggerAfterEach processes each trigger before executing the next.
+// Starting with the first subtrigger, ready when the _current_ subtrigger
+// is ready. After output, advances the current trigger by one.
+type TriggerAfterEach struct {
+ SubTriggers []Trigger
+}
+
+func (t *TriggerAfterEach) onElement(input triggerInput, state *StateData) {
+ ts := state.getTriggerState(t)
+ if ts.finished {
+ return
+ }
+ // Only update the first unfinished sub trigger.
+ for _, sub := range t.SubTriggers {
+ if state.getTriggerState(sub).finished {
+ continue
+ }
+ sub.onElement(input, state)
+ return
+ }
+}
+
+func (t *TriggerAfterEach) shouldFire(state *StateData) bool {
+ ts := state.getTriggerState(t)
+ if ts.finished {
+ return false
+ }
+ for _, sub := range t.SubTriggers {
+ if state.getTriggerState(sub).finished {
+ continue
+ }
+ return sub.shouldFire(state)
+ }
+ return false
+}
+
+func (t *TriggerAfterEach) onFire(state *StateData) {
+ if !t.shouldFire(state) {
+ return
+ }
+ for _, sub := range t.SubTriggers {
+ if state.getTriggerState(sub).finished {
+ continue
+ }
+ sub.onFire(state)
+ if !state.getTriggerState(sub).finished {
+ return
+ }
+ }
+ triggerClearAndFinish(t, state)
+}
+
+func (t *TriggerAfterEach) reset(state *StateData) {
+ subTriggersReset(t, state, t.SubTriggers)
+}
+
+func (t *TriggerAfterEach) String() string {
+ return fmt.Sprintf("AfterEach[%v]", t.SubTriggers)
+}
+
+// TriggerElementCount triggers when there have been at least the required
number
+// of elements have arrived.
+//
+// TriggerElementCount stores the current element count in it's extra state
field.
+type TriggerElementCount struct {
+ ElementCount int
+}
+
+func (t *TriggerElementCount) onElement(input triggerInput, state *StateData) {
+ ts := state.getTriggerState(t)
+ if ts.finished {
+ return
+ }
+
+ if ts.extra == nil {
+ ts.extra = int(0)
+ }
+ count := ts.extra.(int) + input.newElementCount
+ ts.extra = count
+ state.setTriggerState(t, ts)
+}
+
+func (t *TriggerElementCount) shouldFire(state *StateData) bool {
+ ts := state.getTriggerState(t)
+ if ts.finished {
+ return false
+ }
+ if ts.extra == nil {
+ return false
+ }
+ return ts.extra.(int) >= t.ElementCount
+}
+
+func (t *TriggerElementCount) onFire(state *StateData) {
+ if !t.shouldFire(state) {
+ return
+ }
+ ts := state.getTriggerState(t)
+ ts.finished = true
+ ts.extra = nil
+ state.setTriggerState(t, ts)
+}
+
+func (t *TriggerElementCount) reset(state *StateData) {
+ delete(state.Trigger, t)
+}
+
+func (t *TriggerElementCount) String() string {
+ return fmt.Sprintf("ElementCount[%v]", t.ElementCount)
+}
+
+// TriggerOrFinally is ready whenever either of it's subtriggers fire.
+// Ceases to be ready after the Finally trigger shouldFire.
+type TriggerOrFinally struct {
+ Main Trigger // repeated
+ Finally Trigger // terminates execution.
+}
+
+func (t *TriggerOrFinally) onElement(input triggerInput, state *StateData) {
+ ts := state.getTriggerState(t)
+ if ts.finished {
+ return
+ }
+ t.Main.onElement(input, state)
+ t.Finally.onElement(input, state)
+}
+
+func (t *TriggerOrFinally) shouldFire(state *StateData) bool {
+ ts := state.getTriggerState(t)
+ if ts.finished {
+ return false
+ }
+ return t.Main.shouldFire(state) || t.Finally.shouldFire(state)
}
+
+func (t *TriggerOrFinally) onFire(state *StateData) {
+ if !t.shouldFire(state) {
+ return
+ }
+ if t.Finally.shouldFire(state) {
+ t.Finally.onFire(state)
+ ts := state.getTriggerState(t)
+ ts.finished = true
+ state.setTriggerState(t, ts)
+ } else {
+ t.Main.onFire(state)
+ if state.getTriggerState(t.Main).finished {
+ t.Main.reset(state)
+ }
+ }
+}
+
+func (t *TriggerOrFinally) reset(state *StateData) {
+ t.Main.reset(state)
+ t.Finally.reset(state)
+ delete(state.Trigger, t)
+}
+
+func (t *TriggerOrFinally) String() string {
+ return fmt.Sprintf("OrFinally[Repeat:%v Until:%v]", t.Main, t.Finally)
+}
+
+// TriggerRepeatedly is a composite trigger that will fire whenever the
Repeated trigger is ready.
+// If the Repeated trigger is finished, it's state will be reset.
+type TriggerRepeatedly struct {
+ Repeated Trigger
+}
+
+func (t *TriggerRepeatedly) onElement(input triggerInput, state *StateData) {
+ t.Repeated.onElement(input, state)
+}
+
+func (t *TriggerRepeatedly) shouldFire(state *StateData) bool {
+ return t.Repeated.shouldFire(state)
+}
+
+func (t *TriggerRepeatedly) onFire(state *StateData) {
+ if !t.shouldFire(state) {
+ return
+ }
+ t.Repeated.onFire(state)
+ // If the subtrigger is finished, reset it.
+ if repeatedTs := state.getTriggerState(t.Repeated); repeatedTs.finished
{
+ t.Repeated.reset(state)
+ }
+}
+
+func (t *TriggerRepeatedly) reset(state *StateData) {
+ t.Repeated.reset(state)
+ delete(state.Trigger, t)
+}
+
+func (t *TriggerRepeatedly) String() string {
+ return fmt.Sprintf("Repeat[%v]", t.Repeated)
+}
+
+// TriggerAfterEndOfWindow is a composite trigger that will fire whenever the
+// the early Triggers are ready prior to the end of window, implicitly
repeated.
+// After the end of window, the Late trigger will be implicitly repeated.
+//
+// Uses the extra state field to track if the end of the window has been
reached.
+type TriggerAfterEndOfWindow struct {
+ Early, Late Trigger
+}
+
+func (t *TriggerAfterEndOfWindow) onElement(input triggerInput, state
*StateData) {
+ ts := state.getTriggerState(t)
+ if ts.finished {
+ return
+ }
+ if ts.extra == nil {
+ ts.extra = false
+ }
+ previouslyEndOfWindow := ts.extra.(bool)
+ if !previouslyEndOfWindow && input.endOfWindowReached {
+ // We have transitioned. Clear early state and mark it finished
+ triggerClearAndFinish(t.Early, state)
+ if t.Late == nil {
+ triggerClearAndFinish(t, state)
+ return
+ }
+ }
+ ts.extra = input.endOfWindowReached
+ state.setTriggerState(t, ts)
+
+ if !state.getTriggerState(t.Early).finished {
+ t.Early.onElement(input, state)
+ return
+ } else if t.Late != nil {
+ t.Late.onElement(input, state)
+ }
+}
+
+func (t *TriggerAfterEndOfWindow) shouldFire(state *StateData) bool {
+ ts := state.getTriggerState(t)
+ if ts.finished {
+ return false
+ }
+ if !state.getTriggerState(t.Early).finished {
+ return t.Early.shouldFire(state) || ts.extra.(bool)
+ } else if t.Late == nil {
+ return false
+ }
+ return t.Late.shouldFire(state)
+}
+
+func (t *TriggerAfterEndOfWindow) onFire(state *StateData) {
+ ts := state.getTriggerState(t)
+ if ts.finished {
+ return
+ }
+ if !state.getTriggerState(t.Early).finished {
+ if t.Early.shouldFire(state) {
Review Comment:
At present, in the follow up PR, Prism will always Emit at the end of
window, except for the never trigger, which would be only at the closing time.
It doesn't make sense to me for triggers to be designed or implemented in
such a way to lead to a trivial method of dataloss. That is against the point
of a data processing system.
Is there a benefit to allow for trivial dataloss, vs having the sensible
behavior of always firing at the end of the window?
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
