Skip to content
Closed
Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
33 commits
Select commit Hold shift + click to select a range
64c31e5
Withdrew orphaned Local Activity Records and a retry run's Records.
moetemp Oct 27, 2026
da68244
Merged K8 into K7.
moetemp Oct 30, 2026
d3759a9
Merged K8 into K7.
moetemp Oct 30, 2026
57ecc85
Merged K8 into K7.
moetemp Oct 30, 2026
d08f866
Merged K8 into K7.
moetemp Nov 1, 2026
2c99aeb
Renamed Supersession to Superseded, and StreamRef to StreamHandle.
moetemp Nov 1, 2026
3e2c851
Merged K8 into K7.
moetemp Nov 1, 2026
d6c13aa
Merged K8 into K7.
moetemp Nov 1, 2026
a0f7fcc
Merged K8 into K7.
moetemp Nov 4, 2026
0cb1b60
Fenced a late Local Activity write of a task that never completed.
moetemp Nov 4, 2026
c45cd59
Tracked the task of a Local Activity write with no task facts.
moetemp Nov 4, 2026
7a2b763
Kept a Producer's held epoch after an empty batch withdraws a run.
moetemp Nov 4, 2026
e1a05ec
Kept one meta field per withdrawn task, as the Redis Store does.
moetemp Nov 4, 2026
722920f
Built the run's task facts from task keys only.
moetemp Nov 4, 2026
de49d24
Kept a run's open tasks back to back.
moetemp Nov 4, 2026
b675303
Merged K8 into K7.
moetemp Nov 4, 2026
25aa190
Refused task facts that a Store would drop or misread.
moetemp Nov 4, 2026
12b2730
Merged K8 into K7.
moetemp Nov 5, 2026
267c611
Gave the orphan item a Workflow case with the batch's task.
moetemp Nov 5, 2026
4355fb8
Merged K8 into K7.
moetemp Nov 5, 2026
f46a7cf
Merged K8 into K7.
moetemp Nov 5, 2026
95a8165
Refused a bad task key field as corrupt Store data.
moetemp Nov 5, 2026
b3013b2
Merged K8 into K7.
moetemp Nov 5, 2026
5fa85ee
Called the withdrawal suite's Stream a ref.
moetemp Nov 5, 2026
3cada9d
Merged K8 into K7.
moetemp Nov 5, 2026
3e1629a
Merged K8 into K7.
moetemp Nov 5, 2026
5a387a5
Merged K8 into K7.
moetemp Nov 5, 2026
4618eec
Read a write's task facts from its StreamWriter.
moetemp Nov 5, 2026
34a1e7f
Merged K8 into K7.
moetemp Nov 5, 2026
aa0a92c
Merged K8 into K7.
moetemp Nov 5, 2026
52816a4
Merged K8 into K7.
moetemp Nov 5, 2026
0d342e4
Put the Producer kind and the writer's withdrawal into variants.
moetemp Nov 7, 2026
1145fe0
Read the Producer kind and the writer's withdrawal through the oneofs…
moetemp Nov 7, 2026
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
102 changes: 63 additions & 39 deletions crates/protos/protos/api_upstream/temporal/api/stream/v1/message.proto
Original file line number Diff line number Diff line change
Expand Up @@ -29,27 +29,6 @@ message StreamRef {
string backend = 3;
}

// Experimental. Might change in the future.
// The ordered attempt of a Producer. A higher epoch supersedes the Records of a lower epoch of the
// same Producer. The case is the kind of the Producer. It decides how the Store orders the epochs,
// and which earlier Records a higher epoch withdraws.
message ProducerEpoch {
oneof kind {
// A Workflow run, which writes in its Workflow Task. A higher epoch withdraws the Records
// of the run's Workflow Tasks that never completed, also those of their Local Activities.
WorkflowTaskEpoch workflow = 1;
// A Local Activity. A higher epoch withdraws the Producer's lower epochs.
LocalActivityEpoch local_activity = 2;
// An Activity of a Workflow. A higher epoch withdraws the Producer's lower epochs.
ActivityEpoch activity = 3;
// A Standalone Activity. A higher epoch withdraws the Producer's lower epochs.
ActivityEpoch standalone_activity = 4;
// A client. A higher epoch only fences the lower ones, since a restarted client's earlier
// Records are real messages, not the Records of a failed attempt.
ClientEpoch client = 5;
}
}

// Experimental. Might change in the future.
// One Workflow Task attempt of a run. It orders by started_event_id, then attempt, then
// started_time.
Expand Down Expand Up @@ -86,32 +65,78 @@ message ClientEpoch {
}

// Experimental. Might change in the future.
// The code that wrote a batch. Its id stays the same across attempts, and its epoch orders them.
// A Workflow run Producer. It writes in its Workflow Task. A higher epoch withdraws the Records of
// the run's Workflow Tasks that never completed, with those of their Local Activities.
message WorkflowProducer {
// The first execution run id of the run chain.
string run_id = 1;
WorkflowTaskEpoch epoch = 2;
}

// Experimental. Might change in the future.
// A Local Activity Producer. A higher epoch withdraws the Producer's lower epochs.
message LocalActivityProducer {
string run_id = 1;
LocalActivityEpoch epoch = 2;
}

// Experimental. Might change in the future.
// An Activity Producer, for an Activity of a Workflow or a Standalone Activity. A higher epoch
// withdraws the Producer's lower epochs.
message ActivityProducer {
string run_id = 1;
ActivityEpoch epoch = 2;
}

// Experimental. Might change in the future.
// A client Producer. A higher epoch only fences the lower ones, since a restarted client's earlier
// Records are real messages, not the Records of a failed attempt. A client has no run.
message ClientProducer {
ClientEpoch epoch = 1;
}

// Experimental. Might change in the future.
// The code that wrote a batch. Its id stays the same across attempts, and the kind's case orders
// its epochs and picks which earlier Records a higher epoch withdraws. A run-based kind carries its
// run, so a client can't.
message Producer {
// Stable across attempts, such as `wf:<run id>` for the Workflow Producer of a run.
string id = 1;
// Required. Its case gives the kind of the Producer.
ProducerEpoch epoch = 2;
// The run of the Producer, so nobody parses the id. Empty for a client.
string run_id = 3;
// Required. The case is the kind of the Producer.
oneof kind {
WorkflowProducer workflow = 2;
LocalActivityProducer local_activity = 3;
ActivityProducer activity = 4;
ActivityProducer standalone_activity = 5;
ClientProducer client = 6;
}
}

// Experimental. Might change in the future.
// The Producer of a write, and which earlier Records the write withdraws. An append, a close and
// a writer carry the same fields, so the Store orders and withdraws each write the same way.
message StreamWriter {
// Required on a write. Its epoch's case picks the withdrawal rule.
// Required on a write. Its Producer's kind picks the withdrawal rule.
Producer producer = 1;
// Workflow and Local Activity only. The run's last completed Workflow Task, from a Local
// Activity's start. Its started_event_id is 0 before the first. The write withdraws the run's
// Local Activity Records whose Workflow Task never completed.
WorkflowTaskEpoch previous_completed_task = 2;
// Workflow and Local Activity only. The started event ids of the run's failed or timed out
// Workflow Tasks. The write withdraws the run's Local Activity Records of these tasks.
repeated int64 failed_tasks = 3;
// Retry run only. The previous run, on the retry run's first write. Every Record whose
// Producer carries it is withdrawn.
string withdraw_run = 4;
// Set only by a Workflow or Local Activity write, since only a run withdraws orphaned Local
// Activity Records or a previous run. Unset for an Activity or client write.
RunWithdrawal withdrawal = 2;
}

// Experimental. Might change in the future.
// What a Workflow or Local Activity write withdraws from its run. One write can carry both the task
// facts and a run to withdraw, so these are fields of one message and not a oneof.
message RunWithdrawal {
// The run's last completed Workflow Task, from a Local Activity's start. Its started_event_id
// is 0 before the first. The write withdraws the run's Local Activity Records whose Workflow
// Task never completed.
WorkflowTaskEpoch previous_completed_task = 1;
// The started event ids of the run's failed or timed out Workflow Tasks. The write withdraws
// the run's Local Activity Records of these tasks.
repeated int64 failed_tasks = 2;
// Retry run only. The previous run, on the retry run's first write. Every Record whose Producer
// carries it is withdrawn.
string withdraw_run = 3;
}

// Experimental. Might change in the future.
Expand All @@ -125,8 +150,7 @@ message AppendRecord {
// batch of a higher epoch, or for a retry run's first write, and a read builds a Superseded item
// from them. A reader drops the earlier Records that the facts match.
message StreamSuperseded {
// The writer of the batch that withdraws. Its Producer's epoch is empty when only
// withdraw_run applies.
// The writer of the batch that withdraws, with its withdrawal facts.
StreamWriter writer = 1;
}

Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -75,7 +75,7 @@ message AppendRecordsRequest {
int64 sequence = 4;
repeated temporal.api.stream.v1.AppendRecord records = 5;
// Only read by the server, on a Workflow Producer's batch. The batch withdraws no Workflow
// Task and keeps no task facts, but the writer's withdraw_run still applies.
// Task and keeps no task facts, but the writer's withdrawal withdraw_run still applies.
bool defer_withdrawal = 6;
// Only read by the server. The server checks the Producer against the Activity's current
// attempt, so nobody else writes as that Activity.
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -33,9 +33,10 @@ message MemoryStreamConfig {
message OpenWriterRequest {
string namespace = 1;
temporal.api.stream.v1.StreamRef ref = 2;
// Each batch sends it as AppendRecords does. For a Local Activity it has the run's task facts,
// so the writes withdraw the run's orphaned Local Activity Records. Its withdraw_run must be
// empty, since only a retry run's first append withdraws a run, and a writer sends many.
// Each batch sends it as AppendRecords does. For a Local Activity its withdrawal has the run's
// task facts, so the writes withdraw the run's orphaned Local Activity Records. The withdrawal's
// withdraw_run must be empty, since only a retry run's first append withdraws a run, and a
// writer sends many.
temporal.api.stream.v1.StreamWriter writer = 3;
// A write sends at most this many Records in one batch. Zero is 1000.
int32 max_batch_size = 4;
Expand Down
127 changes: 81 additions & 46 deletions crates/streams/src/epoch.rs
Original file line number Diff line number Diff line change
Expand Up @@ -7,8 +7,8 @@
use crate::{
StreamError, StreamResult,
proto::{
Producer, ProducerEpoch, StreamRecord, StreamSuperseded, WorkflowTaskEpoch,
producer_epoch::Kind, stream_record::Item,
Producer, StreamRecord, StreamSuperseded, StreamWriter, WorkflowTaskEpoch, producer::Kind,
stream_record::Item,
},
};
use prost_types::Timestamp;
Expand Down Expand Up @@ -42,22 +42,52 @@ impl From<&Kind> for ProducerKind {
}
}

/// Refuses a Producer with no epoch, or with an epoch that has no case.
/// Refuses a Producer with no kind case.
impl TryFrom<&Producer> for ProducerKind {
type Error = StreamError;

fn try_from(producer: &Producer) -> StreamResult<Self> {
let epoch = producer
.epoch
.as_ref()
.ok_or_else(|| StreamError::invalid("The Producer has no epoch"))?;
kind(epoch).ok_or_else(|| StreamError::invalid("The Producer's epoch has no kind"))
kind_of(producer).ok_or_else(|| StreamError::invalid("The Producer has no kind"))
}
}

/// The kind of an epoch's case, `None` when the epoch has no case.
fn kind(epoch: &ProducerEpoch) -> Option<ProducerKind> {
epoch.kind.as_ref().map(ProducerKind::from)
/// The kind case of a Producer, `None` when it has none.
pub fn kind_of(producer: &Producer) -> Option<ProducerKind> {
producer.kind.as_ref().map(ProducerKind::from)
}

/// The run of a Producer. Empty for a client, which has no run.
pub fn producer_run_id(producer: &Producer) -> &str {
match &producer.kind {
Some(Kind::Workflow(w)) => &w.run_id,
Some(Kind::LocalActivity(l)) => &l.run_id,
Some(Kind::Activity(a) | Kind::StandaloneActivity(a)) => &a.run_id,
Some(Kind::Client(_)) | None => "",
}
}

/// The run that a writer withdraws, empty for none.
pub fn writer_withdraw_run(writer: &StreamWriter) -> &str {
writer
.withdrawal
.as_ref()
.map_or("", |w| w.withdraw_run.as_str())
}

/// The started event ids of the run's failed Workflow Tasks that a writer reports.
pub fn writer_failed_tasks(writer: &StreamWriter) -> &[i64] {
writer
.withdrawal
.as_ref()
.map_or(&[], |w| w.failed_tasks.as_slice())
}

/// The run's last completed Workflow Task that a writer reports, `None` for none.
pub fn writer_previous_completed_task(writer: &StreamWriter) -> Option<&WorkflowTaskEpoch> {
writer
.withdrawal
.as_ref()
.and_then(|w| w.previous_completed_task.as_ref())
}

/// The length of a [task_key].
Expand Down Expand Up @@ -92,20 +122,24 @@ pub fn task_key(task: &WorkflowTaskEpoch) -> [u8; TASK_KEY_LEN] {
}

/// The key of an epoch. It is the [task_of] key, then the attempt's scheduled seconds and nanos
/// and the attempt, in 8, 4 and 4 bytes. A field that the case doesn't have is zero, and an epoch
/// and the attempt, in 8, 4 and 4 bytes. A field that the case doesn't have is zero, and a Producer
/// with no case is all zeros.
pub fn epoch_key(epoch: &ProducerEpoch) -> [u8; EPOCH_KEY_LEN] {
let (scheduled, attempt) = match &epoch.kind {
Some(Kind::LocalActivity(local)) => (None, local.attempt),
pub fn epoch_key(kind: Option<&Kind>) -> [u8; EPOCH_KEY_LEN] {
let (scheduled, attempt) = match kind {
Some(Kind::LocalActivity(local)) => (None, local.epoch.map_or(0, |e| e.attempt)),
Some(Kind::Activity(activity) | Kind::StandaloneActivity(activity)) => {
(activity.attempt_scheduled_time.as_ref(), activity.attempt)
let epoch = activity.epoch;
(
epoch.and_then(|e| e.attempt_scheduled_time),
epoch.map_or(0, |e| e.attempt),
)
}
Some(Kind::Client(client)) => (None, client.attempt),
Some(Kind::Client(client)) => (None, client.epoch.map_or(0, |e| e.attempt)),
Some(Kind::Workflow(_)) | None => (None, 0),
};
let (seconds, nanos) = time_fields(scheduled);
let (seconds, nanos) = time_fields(scheduled.as_ref());
let mut key = [0; EPOCH_KEY_LEN];
key[..TASK_KEY_LEN].copy_from_slice(&task_key(&task_of(epoch)));
key[..TASK_KEY_LEN].copy_from_slice(&task_key(&task_of(kind)));
key[24..32].copy_from_slice(&be64(seconds));
key[32..36].copy_from_slice(&be32(nanos));
key[36..].copy_from_slice(&be32(attempt));
Expand All @@ -114,16 +148,16 @@ pub fn epoch_key(epoch: &ProducerEpoch) -> [u8; EPOCH_KEY_LEN] {

/// The Workflow Task of an epoch. It is a Workflow's own task, or the task that dispatched a Local
/// Activity, and zero for the other kinds.
pub fn task_of(epoch: &ProducerEpoch) -> WorkflowTaskEpoch {
match &epoch.kind {
Some(Kind::Workflow(task)) => *task,
Some(Kind::LocalActivity(local)) => local.task.unwrap_or_default(),
pub fn task_of(kind: Option<&Kind>) -> WorkflowTaskEpoch {
match kind {
Some(Kind::Workflow(w)) => w.epoch.unwrap_or_default(),
Some(Kind::LocalActivity(local)) => local.epoch.and_then(|e| e.task).unwrap_or_default(),
_ => WorkflowTaskEpoch::default(),
}
}

/// Orders two epochs field by field.
pub fn compare_epochs(a: &ProducerEpoch, b: &ProducerEpoch) -> Ordering {
pub fn compare_epochs(a: Option<&Kind>, b: Option<&Kind>) -> Ordering {
epoch_key(a).cmp(&epoch_key(b))
}

Expand All @@ -150,66 +184,67 @@ pub fn withdraws(superseded: &StreamSuperseded, record: &StreamRecord) -> bool {
let Some(producer) = data.producer.as_ref() else {
return false;
};
if !writer.withdraw_run.is_empty() && producer.run_id == writer.withdraw_run {
let withdraw_run = writer_withdraw_run(writer);
if !withdraw_run.is_empty() && producer_run_id(producer) == withdraw_run {
return true;
}
let Some(withdrawing) = writer.producer.as_ref() else {
return false;
};
let (Some(epoch), Some(old)) = (withdrawing.epoch.as_ref(), producer.epoch.as_ref()) else {
let (new, old) = (withdrawing.kind.as_ref(), producer.kind.as_ref());
let (Some(new_kind), Some(old_kind)) = (new, old) else {
return false;
};
match kind(epoch) {
None | Some(ProducerKind::Client) => return false,
Some(ProducerKind::Workflow) => {}
match ProducerKind::from(new_kind) {
ProducerKind::Client => return false,
ProducerKind::Workflow => {}
same => {
return kind(old) == same
return ProducerKind::from(old_kind) == same
&& producer.id == withdrawing.id
&& compare_epochs(old, epoch) == Ordering::Less;
&& compare_epochs(old, new) == Ordering::Less;
}
}
if !of_run(producer, withdrawing) {
return false;
}
let (task, old) = (task_of(epoch), task_of(old));
let started = old.started_event_id;
if writer.failed_tasks.contains(&started) {
let (task, old_task) = (task_of(new), task_of(old));
let started = old_task.started_event_id;
if writer_failed_tasks(writer).contains(&started) {
return true;
}
let completed = writer.previous_completed_task.as_ref();
let completed = writer_previous_completed_task(writer);
let after = completed.map_or(0, |c| c.started_event_id);
if after < started && started < task.started_event_id {
return true;
}
let lower_at = |at: &WorkflowTaskEpoch| {
started == at.started_event_id && compare_tasks(&old, at) == Ordering::Less
started == at.started_event_id && compare_tasks(&old_task, at) == Ordering::Less
};
lower_at(&task) || completed.is_some_and(lower_at)
}

/// Whether a Producer is the Workflow run that withdraws, or one of its Local Activities.
fn of_run(producer: &Producer, withdrawing: &Producer) -> bool {
match producer.epoch.as_ref().and_then(kind) {
match kind_of(producer) {
Some(ProducerKind::Workflow) => producer.id == withdrawing.id,
Some(ProducerKind::LocalActivity) => producer.run_id == withdrawing.run_id,
Some(ProducerKind::LocalActivity) => {
producer_run_id(producer) == producer_run_id(withdrawing)
}
_ => false,
}
}

#[cfg(test)]
mod tests {
use super::*;
use crate::proto::ClientEpoch;
use crate::proto::{ClientEpoch, ClientProducer, producer::Kind};

// Without the clamp, the bytes of a negative field would outrank every real epoch.
#[test]
fn a_negative_field_compares_as_zero() {
let negative = ProducerEpoch {
kind: Some(Kind::Client(ClientEpoch { attempt: -1 })),
};
assert_eq!(
compare_epochs(&negative, &ProducerEpoch::default()),
Ordering::Equal
);
let negative = Kind::Client(ClientProducer {
epoch: Some(ClientEpoch { attempt: -1 }),
});
assert_eq!(compare_epochs(Some(&negative), None), Ordering::Equal);
}
}
Loading