@temporalio/core-bridge 1.5.2 → 1.6.0

package/sdk-core/core/src/worker/workflow/managed_run.rs

@@ -1,188 +1,228 @@
 #[cfg(test)]
 mod managed_wf_test;
 
+#[cfg(test)]
+pub(crate) use managed_wf_test::ManagedWFFunc;
+
 use crate::{
+    abstractions::dbg_panic,
+    protosext::WorkflowActivationExt,
     worker::{
         workflow::{
-            machines::WorkflowMachines, ActivationAction, ActivationCompleteOutcome, HistoryUpdate,
-            LocalResolution, NewIncomingWFT, OutgoingServerCommands, RequestEvictMsg, RunActions,
-            RunActivationCompletion, RunUpdateResponse, ServerCommandsWithWorkflowInfo, WFCommand,
-            WorkflowBridge,
+            history_update::HistoryPaginator, machines::WorkflowMachines, ActivationAction,
+            ActivationCompleteOutcome, ActivationCompleteResult, ActivationOrAuto,
+            EvictionRequestResult, FailedActivationWFTReport, HeartbeatTimeoutMsg, HistoryUpdate,
+            LocalActivityRequestSink, LocalResolution, NextPageReq, OutgoingServerCommands,
+            OutstandingActivation, OutstandingTask, PermittedWFT, RequestEvictMsg,
+            ServerCommandsWithWorkflowInfo, WFCommand, WFMachinesError, WFTReportStatus,
+            WorkflowBridge, WorkflowTaskInfo, WFT_HEARTBEAT_TIMEOUT_FRACTION,
         },
-        LocalActRequest,
+        LocalActRequest, LEGACY_QUERY_ID,
     },
     MetricsContext,
 };
-use futures::{stream, StreamExt};
+use futures_util::future::AbortHandle;
 use std::{
+    collections::HashSet,
     ops::Add,
-    sync::mpsc::Sender,
+    sync::{mpsc::Sender, Arc},
     time::{Duration, Instant},
 };
-use temporal_sdk_core_api::errors::WFMachinesError;
-use temporal_sdk_core_protos::coresdk::{
-    workflow_activation::{RemoveFromCache, WorkflowActivation},
-    workflow_commands::QueryResult,
-};
-use tokio::{
-    sync::{
-        mpsc::{unbounded_channel, UnboundedReceiver, UnboundedSender},
-        oneshot,
+use temporal_sdk_core_protos::{
+    coresdk::{
+        workflow_activation::{
+            create_evict_activation, query_to_job, remove_from_cache::EvictionReason,
+            workflow_activation_job, RemoveFromCache, WorkflowActivation,
+        },
+        workflow_commands::QueryResult,
+        workflow_completion,
     },
-    task,
-    task::JoinHandle,
+    temporal::api::{enums::v1::WorkflowTaskFailedCause, failure::v1::Failure},
+    TaskToken,
 };
-use tokio_stream::wrappers::UnboundedReceiverStream;
+use tokio::sync::oneshot;
 use tracing::Span;
-use tracing_futures::Instrument;
-
-use crate::worker::workflow::{
-    ActivationCompleteResult, ActivationOrAuto, FailRunUpdate, FulfillableActivationComplete,
-    GoodRunUpdate, LocalActivityRequestSink, RunAction, RunUpdateResponseKind,
-};
-use temporal_sdk_core_protos::TaskToken;
-
-use crate::abstractions::dbg_panic;
-#[cfg(test)]
-pub(crate) use managed_wf_test::ManagedWFFunc;
 
 type Result<T, E = WFMachinesError> = std::result::Result<T, E>;
-/// What percentage of a WFT timeout we are willing to wait before sending a WFT heartbeat when
-/// necessary.
-const WFT_HEARTBEAT_TIMEOUT_FRACTION: f32 = 0.8;
+pub(super) type RunUpdateAct = Option<ActivationOrAuto>;
 
+/// Manages access to a specific workflow run. Everything inside is entirely synchronous and should
+/// remain that way.
+#[derive(derive_more::DebugCustom)]
+#[debug(
+    fmt = "ManagedRun {{ wft: {:?}, activation: {:?}, buffered_resp: {:?} \
+           trying_to_evict: {} }}",
+    wft,
+    activation,
+    buffered_resp,
+    "trying_to_evict.is_some()"
+)]
 pub(super) struct ManagedRun {
     wfm: WorkflowManager,
-    update_tx: UnboundedSender<RunUpdateResponse>,
-    local_activity_request_sink: LocalActivityRequestSink,
+    /// Called when the machines need to produce local activity requests. This can't be lifted up
+    /// easily as return values, because sometimes local activity requests trigger immediate
+    /// resolutions (ex: too many attempts). Thus lifting it up creates a lot of unneeded complexity
+    /// pushing things out and then directly back in. The downside is this is the only "impure" part
+    /// of the in/out nature of workflow state management. If there's ever a sensible way to lift it
+    /// up, that'd be nice.
+    local_activity_request_sink: Arc<dyn LocalActivityRequestSink>,
+    /// Set if the run is currently waiting on the execution of some local activities.
     waiting_on_la: Option<WaitingOnLAs>,
-    // Is set to true if the machines encounter an error and the only subsequent thing we should
-    // do is be evicted.
+    /// Is set to true if the machines encounter an error and the only subsequent thing we should
+    /// do is be evicted.
     am_broken: bool,
-}
+    /// If set, the WFT this run is currently/will be processing.
+    wft: Option<OutstandingTask>,
+    /// An outstanding activation to lang
+    activation: Option<OutstandingActivation>,
+    /// If set, it indicates there is a buffered poll response from the server that applies to this
+    /// run. This can happen when lang takes too long to complete a task and the task times out, for
+    /// example. Upon next completion, the buffered response will be removed and can be made ready
+    /// to be returned from polling
+    buffered_resp: Option<PermittedWFT>,
+    /// Is set if an eviction has been requested for this run
+    trying_to_evict: Option<RequestEvictMsg>,
 
-/// If an activation completion needed to wait on LA completions (or heartbeat timeout) we use
-/// this struct to store the data we need to finish the completion once that has happened
-struct WaitingOnLAs {
-    wft_timeout: Duration,
-    /// If set, we are waiting for LAs to complete as part of a just-finished workflow activation.
-    /// If unset, we already had a heartbeat timeout and got a new WFT without any new work while
-    /// there are still incomplete LAs.
-    completion_dat: Option<(
-        CompletionDataForWFT,
-        oneshot::Sender<ActivationCompleteResult>,
-    )>,
-    hb_chan: UnboundedSender<Span>,
-    heartbeat_timeout_task: JoinHandle<()>,
-}
-
-#[derive(Debug)]
-struct CompletionDataForWFT {
-    task_token: TaskToken,
-    query_responses: Vec<QueryResult>,
-    has_pending_query: bool,
-    activation_was_only_eviction: bool,
+    /// We track if we have recorded useful debugging values onto a certain span yet, to overcome
+    /// duplicating field values. Remove this once https://github.com/tokio-rs/tracing/issues/2334
+    /// is fixed.
+    recorded_span_ids: HashSet<tracing::Id>,
+    metrics: MetricsContext,
+    /// We store the paginator used for our own run's history fetching
+    paginator: Option<HistoryPaginator>,
+    completion_waiting_on_page_fetch: Option<RunActivationCompletion>,
 }
-
 impl ManagedRun {
+    #[allow(clippy::too_many_arguments)] // Ok with this here. Nothing reusable to extract.
     pub(super) fn new(
-        wfm: WorkflowManager,
-        update_tx: UnboundedSender<RunUpdateResponse>,
-        local_activity_request_sink: LocalActivityRequestSink,
+        history_update: HistoryUpdate,
+        namespace: String,
+        workflow_id: String,
+        workflow_type: String,
+        run_id: String,
+        local_activity_request_sink: Arc<dyn LocalActivityRequestSink>,
+        metrics: MetricsContext,
     ) -> Self {
+        let wfm = WorkflowManager::new(
+            history_update,
+            namespace,
+            workflow_id,
+            workflow_type,
+            run_id,
+            metrics.clone(),
+        );
         Self {
             wfm,
-            update_tx,
             local_activity_request_sink,
             waiting_on_la: None,
             am_broken: false,
+            wft: None,
+            activation: None,
+            buffered_resp: None,
+            trying_to_evict: None,
+            recorded_span_ids: Default::default(),
+            metrics,
+            paginator: None,
+            completion_waiting_on_page_fetch: None,
         }
     }
 
-    pub(super) async fn run(self, run_actions_rx: UnboundedReceiver<RunAction>) {
-        let (heartbeat_tx, heartbeat_rx) = unbounded_channel();
-        stream::select(
-            UnboundedReceiverStream::new(run_actions_rx),
-            UnboundedReceiverStream::new(heartbeat_rx).map(|trace_span| RunAction {
-                action: RunActions::HeartbeatTimeout,
-                trace_span,
-            }),
-        )
-        .fold((self, heartbeat_tx), |(mut me, heartbeat_tx), action| {
-            let span = action.trace_span;
-            let action = action.action;
-            let mut no_wft = false;
-            async move {
-                let res = match action {
-                    RunActions::NewIncomingWFT(wft) => me
-                        .incoming_wft(wft)
-                        .await
-                        .map(RunActionOutcome::AfterNewWFT),
-                    RunActions::ActivationCompletion(completion) => me
-                        .completion(completion, &heartbeat_tx)
-                        .await
-                        .map(RunActionOutcome::AfterCompletion),
-                    RunActions::CheckMoreWork {
-                        want_to_evict,
-                        has_pending_queries,
-                        has_wft,
-                    } => {
-                        if !has_wft {
-                            no_wft = true;
-                        }
-                        me.check_more_work(want_to_evict, has_pending_queries, has_wft)
-                            .await
-                            .map(RunActionOutcome::AfterCheckWork)
-                    }
-                    RunActions::LocalResolution(r) => me
-                        .local_resolution(r)
-                        .await
-                        .map(RunActionOutcome::AfterLocalResolution),
-                    RunActions::HeartbeatTimeout => {
-                        let maybe_act = if me.heartbeat_timeout() {
-                            Some(ActivationOrAuto::Autocomplete {
-                                run_id: me.wfm.machines.run_id.clone(),
-                            })
-                        } else {
-                            None
-                        };
-                        Ok(RunActionOutcome::AfterHeartbeatTimeout(maybe_act))
-                    }
-                };
-                match res {
-                    Ok(outcome) => {
-                        me.send_update_response(outcome, no_wft);
-                    }
-                    Err(e) => {
-                        error!(error=?e, "Error in run machines");
-                        me.am_broken = true;
-                        me.update_tx
-                            .send(RunUpdateResponse {
-                                kind: RunUpdateResponseKind::Fail(FailRunUpdate {
-                                    run_id: me.wfm.machines.run_id.clone(),
-                                    err: e.source,
-                                    completion_resp: e.complete_resp_chan,
-                                }),
-                                span: Span::current(),
-                            })
-                            .expect("Machine can send update");
-                    }
-                }
-                (me, heartbeat_tx)
-            }
-            .instrument(span)
-        })
-        .await;
+    /// Returns true if there are pending jobs that need to be sent to lang.
+    pub(super) fn more_pending_work(&self) -> bool {
+        // We don't want to consider there to be more local-only work to be done if there is
+        // no workflow task associated with the run right now. This can happen if, ex, we
+        // complete a local activity while waiting for server to send us the next WFT.
+        // Activating lang would be harmful at this stage, as there might be work returned
+        // in that next WFT which should be part of the next activation.
+        self.wft.is_some() && self.wfm.machines.has_pending_jobs()
+    }
+
+    pub(super) fn have_seen_terminal_event(&self) -> bool {
+        self.wfm.machines.have_seen_terminal_event
+    }
+
+    /// Returns a ref to info about the currently tracked workflow task, if any.
+    pub(super) fn wft(&self) -> Option<&OutstandingTask> {
+        self.wft.as_ref()
+    }
+
+    /// Returns a ref to info about the currently tracked workflow activation, if any.
+    pub(super) fn activation(&self) -> Option<&OutstandingActivation> {
+        self.activation.as_ref()
+    }
+
+    /// Returns true if this run has already been told it will be evicted.
+    pub(super) fn is_trying_to_evict(&self) -> bool {
+        self.trying_to_evict.is_some()
     }
 
-    async fn incoming_wft(
+    /// Called whenever a new workflow task is obtained for this run
+    pub(super) fn incoming_wft(&mut self, pwft: PermittedWFT) -> RunUpdateAct {
+        let res = self._incoming_wft(pwft);
+        self.update_to_acts(res.map(Into::into), true)
+    }
+
+    fn _incoming_wft(
         &mut self,
-        wft: NewIncomingWFT,
+        pwft: PermittedWFT,
     ) -> Result<Option<ActivationOrAuto>, RunUpdateErr> {
-        let activation = if let Some(h) = wft.history_update {
-            self.wfm.feed_history_from_server(h).await?
+        if self.wft.is_some() {
+            dbg_panic!("Trying to send a new WFT for a run which already has one!");
+        }
+        let start_time = Instant::now();
+
+        let work = pwft.work;
+        let did_miss_cache = !work.is_incremental() || !work.update.is_real();
+        debug!(
+            run_id = %work.execution.run_id,
+            task_token = %&work.task_token,
+            update = ?work.update,
+            has_legacy_query = %work.legacy_query.is_some(),
+            attempt = %work.attempt,
+            "Applying new workflow task from server"
+        );
+        let wft_info = WorkflowTaskInfo {
+            attempt: work.attempt,
+            task_token: work.task_token,
+            wf_id: work.execution.workflow_id.clone(),
+        };
+
+        let legacy_query_from_poll = work
+            .legacy_query
+            .map(|q| query_to_job(LEGACY_QUERY_ID.to_string(), q));
+
+        let mut pending_queries = work.query_requests;
+        if !pending_queries.is_empty() && legacy_query_from_poll.is_some() {
+            error!(
+                "Server issued both normal and legacy queries. This should not happen. Please \
+                 file a bug report."
+            );
+            return Err(RunUpdateErr {
+                source: WFMachinesError::Fatal(
+                    "Server issued both normal and legacy query".to_string(),
+                ),
+                complete_resp_chan: None,
+            });
+        }
+        if let Some(lq) = legacy_query_from_poll {
+            pending_queries.push(lq);
+        }
+
+        self.paginator = Some(pwft.paginator);
+        self.wft = Some(OutstandingTask {
+            info: wft_info,
+            hit_cache: !did_miss_cache,
+            pending_queries,
+            start_time,
+            permit: pwft.permit,
+        });
+
+        // The update field is only populated in the event we hit the cache
+        let activation = if work.update.is_real() {
+            self.metrics.sticky_cache_hit();
+            self.wfm.feed_history_from_server(work.update)?
         } else {
-            let r = self.wfm.get_next_activation().await?;
+            let r = self.wfm.get_next_activation()?;
             if r.jobs.is_empty() {
                 return Err(RunUpdateErr {
                     source: WFMachinesError::Fatal(format!(
@@ -197,16 +237,17 @@ impl ManagedRun {
 
         if activation.jobs.is_empty() {
             if self.wfm.machines.outstanding_local_activity_count() > 0 {
-                // If the activation has no jobs but there are outstanding LAs, we need to restart the
-                // WFT heartbeat.
+                // If the activation has no jobs but there are outstanding LAs, we need to restart
+                // the WFT heartbeat.
                 if let Some(ref mut lawait) = self.waiting_on_la {
                     if lawait.completion_dat.is_some() {
                         panic!("Should not have completion dat when getting new wft & empty jobs")
                     }
-                    lawait.heartbeat_timeout_task.abort();
-                    lawait.heartbeat_timeout_task = start_heartbeat_timeout_task(
-                        lawait.hb_chan.clone(),
-                        wft.start_time,
+                    lawait.hb_timeout_handle.abort();
+                    lawait.hb_timeout_handle = sink_heartbeat_timeout_start(
+                        self.wfm.machines.run_id.clone(),
+                        self.local_activity_request_sink.as_ref(),
+                        start_time,
                         lawait.wft_timeout,
                     );
                     // No activation needs to be sent to lang. We just need to wait for another
@@ -228,41 +269,332 @@ impl ManagedRun {
         Ok(Some(ActivationOrAuto::LangActivation(activation)))
     }
 
-    async fn completion(
+    /// Deletes the currently tracked WFT & records latency metrics. Should be called after it has
+    /// been responded to (server has been told). Returns the WFT if there was one.
+    pub(super) fn mark_wft_complete(
         &mut self,
-        mut completion: RunActivationCompletion,
-        heartbeat_tx: &UnboundedSender<Span>,
-    ) -> Result<Option<FulfillableActivationComplete>, RunUpdateErr> {
-        let resp_chan = completion
-            .resp_chan
-            .take()
-            .expect("Completion response channel must be populated");
-
-        let outcome = async move {
-            // Send commands from lang into the machines then check if the workflow run
-            // needs another activation and mark it if so
-            self.wfm.push_commands(completion.commands).await?;
-            // Don't bother applying the next task if we're evicting at the end of
-            // this activation
-            if !completion.activation_was_eviction {
-                self.wfm.apply_next_task_if_ready().await?;
+        report_status: WFTReportStatus,
+    ) -> Option<OutstandingTask> {
+        debug!("Marking WFT completed");
+        let retme = self.wft.take();
+
+        // Only record latency metrics if we genuinely reported to server
+        if matches!(report_status, WFTReportStatus::Reported) {
+            if let Some(ot) = &retme {
+                self.metrics.wf_task_latency(ot.start_time.elapsed());
             }
-            let new_local_acts = self.wfm.drain_queued_local_activities();
+        }
+
+        retme
+    }
+
+    /// Checks if any further activations need to go out for this run and produces them if so.
+    pub(super) fn check_more_activations(&mut self) -> RunUpdateAct {
+        let res = self._check_more_activations();
+        self.update_to_acts(res.map(Into::into), false)
+    }
+
+    fn _check_more_activations(&mut self) -> Result<Option<ActivationOrAuto>, RunUpdateErr> {
+        // No point in checking for more activations if there's already an outstanding activation.
+        if self.activation.is_some() {
+            return Ok(None);
+        }
+        // In the event it's time to evict this run, cancel any outstanding LAs
+        if self.trying_to_evict.is_some() {
+            self.sink_la_requests(vec![LocalActRequest::CancelAllInRun(
+                self.wfm.machines.run_id.clone(),
+            )])?;
+        }
+
+        if self.wft.is_none() {
+            // It doesn't make sense to do workflow work unless we have a WFT
+            return Ok(None);
+        }
+
+        if self.wfm.machines.has_pending_jobs() && !self.am_broken {
+            Ok(Some(ActivationOrAuto::LangActivation(
+                self.wfm.get_next_activation()?,
+            )))
+        } else {
+            if !self.am_broken {
+                let has_pending_queries = self
+                    .wft
+                    .as_ref()
+                    .map(|wft| !wft.pending_queries.is_empty())
+                    .unwrap_or_default();
+                if has_pending_queries {
+                    return Ok(Some(ActivationOrAuto::ReadyForQueries(
+                        self.wfm.machines.get_wf_activation(),
+                    )));
+                }
+            }
+            if let Some(wte) = self.trying_to_evict.clone() {
+                let mut act = self.wfm.machines.get_wf_activation();
+                // No other jobs make any sense to send if we encountered an error.
+                if self.am_broken {
+                    act.jobs = vec![];
+                }
+                act.append_evict_job(RemoveFromCache {
+                    message: wte.message,
+                    reason: wte.reason as i32,
+                });
+                Ok(Some(ActivationOrAuto::LangActivation(act)))
+            } else {
+                Ok(None)
+            }
+        }
+    }
 
-            let immediate_resolutions = (self.local_activity_request_sink)(new_local_acts);
-            for resolution in immediate_resolutions {
-                self.wfm
-                    .notify_of_local_result(LocalResolution::LocalActivity(resolution))?;
+    /// Called whenever lang successfully completes a workflow activation. Commands produced by the
+    /// activation are passed in. `resp_chan` will be used to unblock the completion call when
+    /// everything we need to do to fulfill it has happened.
+    ///
+    /// Can return an error in the event that another page of history needs to be fetched before
+    /// the completion can proceed.
+    pub(super) fn successful_completion(
+        &mut self,
+        mut commands: Vec<WFCommand>,
+        resp_chan: Option<oneshot::Sender<ActivationCompleteResult>>,
+    ) -> Result<RunUpdateAct, NextPageReq> {
+        let activation_was_only_eviction = self.activation_has_only_eviction();
+        let (task_token, has_pending_query, start_time) = if let Some(entry) = self.wft.as_ref() {
+            (
+                entry.info.task_token.clone(),
+                !entry.pending_queries.is_empty(),
+                entry.start_time,
+            )
+        } else {
+            if !activation_was_only_eviction {
+                // Not an error if this was an eviction, since it's normal to issue eviction
+                // activations without an associated workflow task in that case.
+                dbg_panic!(
+                    "Attempted to complete activation for run {} without associated workflow task",
+                    self.run_id()
+                );
             }
+            self.reply_to_complete(ActivationCompleteOutcome::DoNothing, resp_chan);
+            return Ok(None);
+        };
 
-            let data = CompletionDataForWFT {
-                task_token: completion.task_token,
-                query_responses: completion.query_responses,
-                has_pending_query: completion.has_pending_query,
-                activation_was_only_eviction: completion.activation_was_only_eviction,
+        // If the only command from the activation is a legacy query response, that means we need
+        // to respond differently than a typical activation.
+        if matches!(&commands.as_slice(),
+                    &[WFCommand::QueryResponse(qr)] if qr.query_id == LEGACY_QUERY_ID)
+        {
+            let qr = match commands.remove(0) {
+                WFCommand::QueryResponse(qr) => qr,
+                _ => unreachable!("We just verified this is the only command"),
             };
+            self.reply_to_complete(
+                ActivationCompleteOutcome::ReportWFTSuccess(ServerCommandsWithWorkflowInfo {
+                    task_token,
+                    action: ActivationAction::RespondLegacyQuery {
+                        result: Box::new(qr),
+                    },
+                }),
+                resp_chan,
+            );
+            Ok(None)
+        } else {
+            // First strip out query responses from other commands that actually affect machines
+            // Would be prettier with `drain_filter`
+            let mut i = 0;
+            let mut query_responses = vec![];
+            while i < commands.len() {
+                if matches!(commands[i], WFCommand::QueryResponse(_)) {
+                    if let WFCommand::QueryResponse(qr) = commands.remove(i) {
+                        query_responses.push(qr);
+                    }
+                } else {
+                    i += 1;
+                }
+            }
+
+            if activation_was_only_eviction && !commands.is_empty() {
+                dbg_panic!("Reply to an eviction only containing an eviction included commands");
+            }
+
+            let rac = RunActivationCompletion {
+                task_token,
+                start_time,
+                commands,
+                activation_was_eviction: self.activation_has_eviction(),
+                activation_was_only_eviction,
+                has_pending_query,
+                query_responses,
+                resp_chan,
+            };
+
+            // Verify we can actually apply the next workflow task, which will happen as part of
+            // applying the completion to machines. If we can't, return early indicating we need
+            // to fetch a page.
+            if !self.wfm.ready_to_apply_next_wft() {
+                return if let Some(paginator) = self.paginator.take() {
+                    debug!("Need to fetch a history page before next WFT can be applied");
+                    self.completion_waiting_on_page_fetch = Some(rac);
+                    Err(NextPageReq {
+                        paginator,
+                        span: Span::current(),
+                    })
+                } else {
+                    Ok(self.update_to_acts(
+                        Err(RunUpdateErr {
+                            source: WFMachinesError::Fatal(
+                                "Run's paginator was absent when attempting to fetch next history \
+                                page. This is a Core SDK bug."
+                                    .to_string(),
+                            ),
+                            complete_resp_chan: rac.resp_chan,
+                        }),
+                        false,
+                    ))
+                };
+            }
+
+            Ok(self.process_completion(rac))
+        }
+    }
+
+    /// Called after the higher-up machinery has fetched more pages of event history needed to apply
+    /// the next workflow task. The history update and paginator used to perform the fetch are
+    /// passed in, with the update being used to apply the task, and the paginator stored to be
+    /// attached with another fetch request if needed.
+    pub(super) fn fetched_page_completion(
+        &mut self,
+        update: HistoryUpdate,
+        paginator: HistoryPaginator,
+    ) -> RunUpdateAct {
+        let res = self._fetched_page_completion(update, paginator);
+        self.update_to_acts(res.map(Into::into), false)
+    }
+    fn _fetched_page_completion(
+        &mut self,
+        update: HistoryUpdate,
+        paginator: HistoryPaginator,
+    ) -> Result<Option<FulfillableActivationComplete>, RunUpdateErr> {
+        self.paginator = Some(paginator);
+        if let Some(d) = self.completion_waiting_on_page_fetch.take() {
+            self._process_completion(d, Some(update))
+        } else {
+            dbg_panic!(
+                "Shouldn't be possible to be applying a next-page-fetch update when \
+                        doing anything other than completing an activation."
+            );
+            Err(RunUpdateErr::from(WFMachinesError::Fatal(
+                "Tried to apply next-page-fetch update to a run that wasn't handling a completion"
+                    .to_string(),
+            )))
+        }
+    }
+
+    /// Called whenever either core lang cannot complete a workflow activation. EX: Nondeterminism
+    /// or user code threw/panicked, respectively. The `cause` and `reason` fields are determined
+    /// inside core always. The `failure` field may come from lang. `resp_chan` will be used to
+    /// unblock the completion call when everything we need to do to fulfill it has happened.
+    pub(super) fn failed_completion(
+        &mut self,
+        cause: WorkflowTaskFailedCause,
+        reason: EvictionReason,
+        failure: workflow_completion::Failure,
+        resp_chan: Option<oneshot::Sender<ActivationCompleteResult>>,
+    ) -> RunUpdateAct {
+        let tt = if let Some(tt) = self.wft.as_ref().map(|t| t.info.task_token.clone()) {
+            tt
+        } else {
+            dbg_panic!(
+                "No workflow task for run id {} found when trying to fail activation",
+                self.run_id()
+            );
+            self.reply_to_complete(ActivationCompleteOutcome::DoNothing, resp_chan);
+            return None;
+        };
+
+        self.metrics.wf_task_failed();
+        let message = format!("Workflow activation completion failed: {:?}", &failure);
+        // Blow up any cached data associated with the workflow
+        let evict_req_outcome = self.request_eviction(RequestEvictMsg {
+            run_id: self.run_id().to_string(),
+            message,
+            reason,
+        });
+        let should_report = match &evict_req_outcome {
+            EvictionRequestResult::EvictionRequested(Some(attempt), _)
+            | EvictionRequestResult::EvictionAlreadyRequested(Some(attempt)) => *attempt <= 1,
+            _ => false,
+        };
+        let rur = evict_req_outcome.into_run_update_resp();
+        // If the outstanding WFT is a legacy query task, report that we need to fail it
+        let outcome = if self.pending_work_is_legacy_query() {
+            ActivationCompleteOutcome::ReportWFTFail(
+                FailedActivationWFTReport::ReportLegacyQueryFailure(tt, failure),
+            )
+        } else if should_report {
+            ActivationCompleteOutcome::ReportWFTFail(FailedActivationWFTReport::Report(
+                tt, cause, failure,
+            ))
+        } else {
+            ActivationCompleteOutcome::WFTFailedDontReport
+        };
+        self.reply_to_complete(outcome, resp_chan);
+        rur
+    }
+
+    /// Delete the currently tracked workflow activation and return it, if any. Should be called
+    /// after the processing of the activation completion, and WFT reporting.
+    pub(super) fn delete_activation(&mut self) -> Option<OutstandingActivation> {
+        self.activation.take()
+    }
+
+    /// Called when local activities resolve
+    pub(super) fn local_resolution(&mut self, res: LocalResolution) -> RunUpdateAct {
+        let res = self._local_resolution(res);
+        self.update_to_acts(res.map(Into::into), false)
+    }
+
+    fn process_completion(&mut self, completion: RunActivationCompletion) -> RunUpdateAct {
+        let res = self._process_completion(completion, None);
+        self.update_to_acts(res.map(Into::into), false)
+    }
+
+    fn _process_completion(
+        &mut self,
+        completion: RunActivationCompletion,
+        new_update: Option<HistoryUpdate>,
+    ) -> Result<Option<FulfillableActivationComplete>, RunUpdateErr> {
+        let data = CompletionDataForWFT {
+            task_token: completion.task_token,
+            query_responses: completion.query_responses,
+            has_pending_query: completion.has_pending_query,
+            activation_was_only_eviction: completion.activation_was_only_eviction,
+        };
+
+        // If this is just bookkeeping after a reply to an only-eviction activation, we can bypass
+        // everything, since there is no reason to continue trying to update machines.
+        if completion.activation_was_only_eviction {
+            return Ok(Some(self.prepare_complete_resp(
+                completion.resp_chan,
+                data,
+                false,
+            )));
+        }
+
+        let outcome = (|| {
+            // Send commands from lang into the machines then check if the workflow run needs
+            // another activation and mark it if so
+            self.wfm.push_commands_and_iterate(completion.commands)?;
+            // If there was a new update included as part of the completion, apply it.
+            if let Some(update) = new_update {
+                self.wfm.feed_history_from_new_page(update)?;
+            }
+            // Don't bother applying the next task if we're evicting at the end of this activation
+            if !completion.activation_was_eviction {
+                self.wfm.apply_next_task_if_ready()?;
+            }
+            let new_local_acts = self.wfm.drain_queued_local_activities();
+            self.sink_la_requests(new_local_acts)?;
+
             if self.wfm.machines.outstanding_local_activity_count() == 0 {
-                Ok((None, data, self))
+                Ok(None)
             } else {
                 let wft_timeout: Duration = self
                     .wfm
@@ -275,28 +607,26 @@ impl ManagedRun {
                                 .to_string(),
                         )
                     })?;
-                let heartbeat_tx = heartbeat_tx.clone();
-                Ok((
-                    Some((heartbeat_tx, completion.start_time, wft_timeout)),
-                    data,
-                    self,
-                ))
+                Ok(Some((completion.start_time, wft_timeout)))
             }
-        }
-        .await;
+        })();
 
         match outcome {
-            Ok((None, data, me)) => Ok(Some(me.prepare_complete_resp(resp_chan, data, false))),
-            Ok((Some((chan, start_t, wft_timeout)), data, me)) => {
-                if let Some(wola) = me.waiting_on_la.as_mut() {
-                    wola.heartbeat_timeout_task.abort();
+            Ok(None) => Ok(Some(self.prepare_complete_resp(
+                completion.resp_chan,
+                data,
+                false,
+            ))),
+            Ok(Some((start_t, wft_timeout))) => {
+                if let Some(wola) = self.waiting_on_la.as_mut() {
+                    wola.hb_timeout_handle.abort();
                 }
-                me.waiting_on_la = Some(WaitingOnLAs {
+                self.waiting_on_la = Some(WaitingOnLAs {
                     wft_timeout,
-                    completion_dat: Some((data, resp_chan)),
-                    hb_chan: chan.clone(),
-                    heartbeat_timeout_task: start_heartbeat_timeout_task(
-                        chan,
+                    completion_dat: Some((data, completion.resp_chan)),
+                    hb_timeout_handle: sink_heartbeat_timeout_start(
+                        self.run_id().to_string(),
+                        self.local_activity_request_sink.as_ref(),
                         start_t,
                         wft_timeout,
                     ),
@@ -305,72 +635,342 @@ impl ManagedRun {
             }
             Err(e) => Err(RunUpdateErr {
                 source: e,
-                complete_resp_chan: Some(resp_chan),
+                complete_resp_chan: completion.resp_chan,
             }),
         }
     }
 
-    async fn check_more_work(
+    fn _local_resolution(
         &mut self,
-        want_to_evict: Option<RequestEvictMsg>,
-        has_pending_queries: bool,
-        has_wft: bool,
-    ) -> Result<Option<ActivationOrAuto>, RunUpdateErr> {
-        if !has_wft {
-            // It doesn't make sense to do work unless we have a WFT
-            return Ok(None);
+        res: LocalResolution,
+    ) -> Result<Option<FulfillableActivationComplete>, RunUpdateErr> {
+        debug!(resolution=?res, "Applying local resolution");
+        self.wfm.notify_of_local_result(res)?;
+        if self.wfm.machines.outstanding_local_activity_count() == 0 {
+            if let Some(mut wait_dat) = self.waiting_on_la.take() {
+                // Cancel the heartbeat timeout
+                wait_dat.hb_timeout_handle.abort();
+                if let Some((completion_dat, resp_chan)) = wait_dat.completion_dat.take() {
+                    return Ok(Some(self.prepare_complete_resp(
+                        resp_chan,
+                        completion_dat,
+                        false,
+                    )));
+                }
+            }
         }
-        if self.wfm.machines.has_pending_jobs() && !self.am_broken {
-            Ok(Some(ActivationOrAuto::LangActivation(
-                self.wfm.get_next_activation().await?,
-            )))
+        Ok(None)
+    }
+
+    pub(super) fn heartbeat_timeout(&mut self) -> RunUpdateAct {
+        let maybe_act = if self._heartbeat_timeout() {
+            Some(ActivationOrAuto::Autocomplete {
+                run_id: self.wfm.machines.run_id.clone(),
+            })
         } else {
-            if has_pending_queries && !self.am_broken {
-                return Ok(Some(ActivationOrAuto::ReadyForQueries(
-                    self.wfm.machines.get_wf_activation(),
-                )));
+            None
+        };
+        self.update_to_acts(Ok(maybe_act).map(Into::into), false)
+    }
+    /// Returns `true` if autocompletion should be issued, which will actually cause us to end up
+    /// in [completion] again, at which point we'll start a new heartbeat timeout, which will
+    /// immediately trigger and thus finish the completion, forcing a new task as it should.
+    fn _heartbeat_timeout(&mut self) -> bool {
+        if let Some(ref mut wait_dat) = self.waiting_on_la {
+            // Cancel the heartbeat timeout
+            wait_dat.hb_timeout_handle.abort();
+            if let Some((completion_dat, resp_chan)) = wait_dat.completion_dat.take() {
+                let compl = self.prepare_complete_resp(resp_chan, completion_dat, true);
+                // Immediately fulfill the completion since the run update will already have
+                // been replied to
+                compl.fulfill();
+            } else {
+                // Auto-reply WFT complete
+                return true;
             }
-            if let Some(wte) = want_to_evict {
-                let mut act = self.wfm.machines.get_wf_activation();
-                // No other jobs make any sense to send if we encountered an error.
-                if self.am_broken {
-                    act.jobs = vec![];
-                }
-                act.append_evict_job(RemoveFromCache {
-                    message: wte.message,
-                    reason: wte.reason as i32,
-                });
-                Ok(Some(ActivationOrAuto::LangActivation(act)))
+        } else {
+            // If a heartbeat timeout happened, we should always have been waiting on LAs
+            dbg_panic!("WFT heartbeat timeout fired but we were not waiting on any LAs");
+        }
+        false
+    }
+
+    /// Returns true if the managed run has any form of pending work
+    /// If `ignore_evicts` is true, pending evictions do not count as pending work.
+    /// If `ignore_buffered` is true, buffered workflow tasks do not count as pending work.
+    pub(super) fn has_any_pending_work(&self, ignore_evicts: bool, ignore_buffered: bool) -> bool {
+        let evict_work = if ignore_evicts {
+            false
+        } else {
+            self.trying_to_evict.is_some()
+        };
+        let act_work = if ignore_evicts {
+            if let Some(ref act) = self.activation {
+                !act.has_only_eviction()
             } else {
-                Ok(None)
+                false
+            }
+        } else {
+            self.activation.is_some()
+        };
+        let buffered = if ignore_buffered {
+            false
+        } else {
+            self.buffered_resp.is_some()
+        };
+        trace!(wft=self.wft.is_some(), buffered=?buffered, more_work=?self.more_pending_work(),
+               act_work, evict_work, "Does run have pending work?");
+        self.wft.is_some() || buffered || self.more_pending_work() || act_work || evict_work
+    }
+
+    /// Stores some work if there is any outstanding WFT or activation for the run. If there was
+    /// not, returns the work back out inside the option.
+    pub(super) fn buffer_wft_if_outstanding_work(
+        &mut self,
+        work: PermittedWFT,
+    ) -> Option<PermittedWFT> {
+        let about_to_issue_evict = self.trying_to_evict.is_some();
+        let has_wft = self.wft().is_some();
+        let has_activation = self.activation().is_some();
+        if has_wft || has_activation || about_to_issue_evict || self.more_pending_work() {
+            debug!(run_id = %self.run_id(),
+                   "Got new WFT for a run with outstanding work, buffering it");
+            self.buffered_resp = Some(work);
+            None
+        } else {
+            Some(work)
+        }
+    }
+
+    /// Returns true if there is a buffered workflow task for this run.
+    pub(super) fn has_buffered_wft(&self) -> bool {
+        self.buffered_resp.is_some()
+    }
+
+    /// Removes and returns the buffered workflow task, if any.
+    pub(super) fn take_buffered_wft(&mut self) -> Option<PermittedWFT> {
+        self.buffered_resp.take()
+    }
+
+    pub(super) fn request_eviction(&mut self, info: RequestEvictMsg) -> EvictionRequestResult {
+        let attempts = self.wft.as_ref().map(|wt| wt.info.attempt);
+
+        // If we were waiting on a page fetch and we're getting evicted because fetching failed,
+        // then make sure we allow the completion to proceed, otherwise we're stuck waiting forever.
+        if self.completion_waiting_on_page_fetch.is_some()
+            && matches!(info.reason, EvictionReason::PaginationOrHistoryFetch)
+        {
+            // We just checked it is some, unwrap OK.
+            let c = self.completion_waiting_on_page_fetch.take().unwrap();
+            let run_upd = self.failed_completion(
+                WorkflowTaskFailedCause::Unspecified,
+                info.reason,
+                Failure::application_failure(info.message, false).into(),
+                c.resp_chan,
+            );
+            return EvictionRequestResult::EvictionRequested(attempts, run_upd);
+        }
+
+        if !self.activation_has_eviction() && self.trying_to_evict.is_none() {
+            debug!(run_id=%info.run_id, reason=%info.message, "Eviction requested");
+            self.trying_to_evict = Some(info);
+            EvictionRequestResult::EvictionRequested(attempts, self.check_more_activations())
+        } else {
+            EvictionRequestResult::EvictionAlreadyRequested(attempts)
+        }
+    }
+
+    pub(super) fn record_span_fields(&mut self, span: &Span) {
+        if let Some(spid) = span.id() {
+            if self.recorded_span_ids.contains(&spid) {
+                return;
+            }
+            self.recorded_span_ids.insert(spid);
+
+            if let Some(wid) = self.wft().map(|wft| &wft.info.wf_id) {
+                span.record("workflow_id", wid.as_str());
+            }
+        }
+    }
+
+    /// Take the result of some update to ourselves and turn it into a return value of zero or more
+    /// actions
+    fn update_to_acts(
+        &mut self,
+        outcome: Result<ActOrFulfill, RunUpdateErr>,
+        in_response_to_wft: bool,
+    ) -> RunUpdateAct {
+        match outcome {
+            Ok(act_or_fulfill) => {
+                let (mut maybe_act, maybe_fulfill) = match act_or_fulfill {
+                    ActOrFulfill::OutgoingAct(a) => (a, None),
+                    ActOrFulfill::FulfillableComplete(c) => (None, c),
+                };
+                // If there's no activation but is pending work, check and possibly generate one
+                if self.more_pending_work() && maybe_act.is_none() {
+                    match self._check_more_activations() {
+                        Ok(oa) => maybe_act = oa,
+                        Err(e) => {
+                            return self.update_to_acts(Err(e), in_response_to_wft);
+                        }
+                    }
+                }
+                let r = match maybe_act {
+                    Some(ActivationOrAuto::LangActivation(mut activation)) => {
+                        if in_response_to_wft {
+                            let wft = self
+                                .wft
+                                .as_mut()
+                                .expect("WFT must exist for run just updated with one");
+                            // If there are in-poll queries, insert jobs for those queries into the
+                            // activation, but only if we hit the cache. If we didn't, those queries
+                            // will need to be dealt with once replay is over
+                            if wft.hit_cache {
+                                put_queries_in_act(&mut activation, wft);
+                            }
+                        }
+
+                        if activation.jobs.is_empty() {
+                            dbg_panic!("Should not send lang activation with no jobs");
+                        }
+                        Some(ActivationOrAuto::LangActivation(activation))
+                    }
+                    Some(ActivationOrAuto::ReadyForQueries(mut act)) => {
+                        if let Some(wft) = self.wft.as_mut() {
+                            put_queries_in_act(&mut act, wft);
+                            Some(ActivationOrAuto::LangActivation(act))
+                        } else {
+                            dbg_panic!("Ready for queries but no WFT!");
+                            None
+                        }
+                    }
+                    a @ Some(ActivationOrAuto::Autocomplete { .. }) => a,
+                    None => {
+                        if let Some(reason) = self.trying_to_evict.as_ref() {
+                            // If we had nothing to do, but we're trying to evict, just do that now
+                            // as long as there's no other outstanding work.
+                            if self.activation.is_none() && !self.more_pending_work() {
+                                let mut evict_act = create_evict_activation(
+                                    self.run_id().to_string(),
+                                    reason.message.clone(),
+                                    reason.reason,
+                                );
+                                evict_act.history_length =
+                                    self.most_recently_processed_event_number() as u32;
+                                Some(ActivationOrAuto::LangActivation(evict_act))
+                            } else {
+                                None
+                            }
+                        } else {
+                            None
+                        }
+                    }
+                };
+                if let Some(f) = maybe_fulfill {
+                    f.fulfill();
+                }
+
+                match r {
+                    // After each run update, check if it's ready to handle any buffered poll
+                    None | Some(ActivationOrAuto::Autocomplete { .. })
+                        if !self.has_any_pending_work(false, true) =>
+                    {
+                        if let Some(bufft) = self.buffered_resp.take() {
+                            self.incoming_wft(bufft)
+                        } else {
+                            None
+                        }
+                    }
+                    Some(r) => {
+                        self.insert_outstanding_activation(&r);
+                        Some(r)
+                    }
+                    None => None,
+                }
+            }
+            Err(fail) => {
+                self.am_broken = true;
+                let rur = if let Some(resp_chan) = fail.complete_resp_chan {
+                    // Automatically fail the workflow task in the event we couldn't update machines
+                    let fail_cause = if matches!(&fail.source, WFMachinesError::Nondeterminism(_)) {
+                        WorkflowTaskFailedCause::NonDeterministicError
+                    } else {
+                        WorkflowTaskFailedCause::Unspecified
+                    };
+                    let wft_fail_str = format!("{:?}", fail.source);
+                    self.failed_completion(
+                        fail_cause,
+                        fail.source.evict_reason(),
+                        Failure::application_failure(wft_fail_str, false).into(),
+                        Some(resp_chan),
+                    )
+                } else {
+                    warn!(error=?fail.source, "Error while updating workflow");
+                    self.request_eviction(RequestEvictMsg {
+                        run_id: self.run_id().to_string(),
+                        message: format!("Error while updating workflow: {:?}", fail.source),
+                        reason: fail.source.evict_reason(),
+                    })
+                    .into_run_update_resp()
+                };
+                rur
+            }
+        }
+    }
+
+    fn insert_outstanding_activation(&mut self, act: &ActivationOrAuto) {
+        let act_type = match &act {
+            ActivationOrAuto::LangActivation(act) | ActivationOrAuto::ReadyForQueries(act) => {
+                if act.is_legacy_query() {
+                    OutstandingActivation::LegacyQuery
+                } else {
+                    OutstandingActivation::Normal {
+                        contains_eviction: act.eviction_index().is_some(),
+                        num_jobs: act.jobs.len(),
+                    }
+                }
             }
+            ActivationOrAuto::Autocomplete { .. } => OutstandingActivation::Autocomplete,
+        };
+        if let Some(old_act) = self.activation {
+            // This is a panic because we have screwed up core logic if this is violated. It must be
+            // upheld.
+            panic!(
+                "Attempted to insert a new outstanding activation {act:?}, but there already was \
+                 one outstanding: {old_act:?}"
+            );
         }
+        self.activation = Some(act_type);
     }
 
     fn prepare_complete_resp(
         &mut self,
-        resp_chan: oneshot::Sender<ActivationCompleteResult>,
+        resp_chan: Option<oneshot::Sender<ActivationCompleteResult>>,
         data: CompletionDataForWFT,
         due_to_heartbeat_timeout: bool,
     ) -> FulfillableActivationComplete {
         let outgoing_cmds = self.wfm.get_server_commands();
+        if data.activation_was_only_eviction && !outgoing_cmds.commands.is_empty() {
+            dbg_panic!(
+                "There should not be any outgoing commands when preparing a completion response \
+                 if the activation was only an eviction. This is an SDK bug."
+            );
+        }
+
         let query_responses = data.query_responses;
         let has_query_responses = !query_responses.is_empty();
         let is_query_playback = data.has_pending_query && !has_query_responses;
         let mut force_new_wft = due_to_heartbeat_timeout;
 
-        // We only actually want to send commands back to the server if there are no more
-        // pending activations and we are caught up on replay. We don't want to complete a wft
-        // if we already saw the final event in the workflow, or if we are playing back for the
-        // express purpose of fulfilling a query. If the activation we sent was *only* an
-        // eviction, and there were no commands produced during iteration, don't send that
+        // We only actually want to send commands back to the server if there are no more pending
+        // activations and we are caught up on replay. We don't want to complete a wft if we already
+        // saw the final event in the workflow, or if we are playing back for the express purpose of
+        // fulfilling a query. If the activation we sent was *only* an eviction, don't send that
         // either.
-        let no_commands_and_evicting =
-            outgoing_cmds.commands.is_empty() && data.activation_was_only_eviction;
         let should_respond = !(self.wfm.machines.has_pending_jobs()
             || outgoing_cmds.replaying
             || is_query_playback
-            || no_commands_and_evicting);
+            || data.activation_was_only_eviction);
         // If there are pending LA resolutions, and we're responding to a query here,
         // we want to make sure to force a new task, as otherwise once we tell lang about
         // the LA resolution there wouldn't be any task to reply to with the result of iterating
@@ -378,17 +978,16 @@ impl ManagedRun {
         if has_query_responses && self.wfm.machines.has_pending_la_resolutions() {
             force_new_wft = true;
         }
-        let to_be_sent = ServerCommandsWithWorkflowInfo {
-            task_token: data.task_token,
-            action: ActivationAction::WftComplete {
-                force_new_wft,
-                commands: outgoing_cmds.commands,
-                query_responses,
-            },
-        };
 
         let outcome = if should_respond || has_query_responses {
-            ActivationCompleteOutcome::ReportWFTSuccess(to_be_sent)
+            ActivationCompleteOutcome::ReportWFTSuccess(ServerCommandsWithWorkflowInfo {
+                task_token: data.task_token,
+                action: ActivationAction::WftComplete {
+                    force_new_wft,
+                    commands: outgoing_cmds.commands,
+                    query_responses,
+                },
+            })
         } else {
             ActivationCompleteOutcome::DoNothing
         };
@@ -401,131 +1000,136 @@ impl ManagedRun {
         }
     }
 
-    async fn local_resolution(
+    /// Pump some local activity requests into the sink, applying any immediate results to the
+    /// workflow machines.
+    fn sink_la_requests(
         &mut self,
-        res: LocalResolution,
-    ) -> Result<Option<FulfillableActivationComplete>, RunUpdateErr> {
-        debug!(resolution=?res, "Applying local resolution");
-        self.wfm.notify_of_local_result(res)?;
-        if self.wfm.machines.outstanding_local_activity_count() == 0 {
-            if let Some(mut wait_dat) = self.waiting_on_la.take() {
-                // Cancel the heartbeat timeout
-                wait_dat.heartbeat_timeout_task.abort();
-                if let Some((completion_dat, resp_chan)) = wait_dat.completion_dat.take() {
-                    return Ok(Some(self.prepare_complete_resp(
-                        resp_chan,
-                        completion_dat,
-                        false,
-                    )));
-                }
-            }
+        new_local_acts: Vec<LocalActRequest>,
+    ) -> Result<(), WFMachinesError> {
+        let immediate_resolutions = self.local_activity_request_sink.sink_reqs(new_local_acts);
+        if !immediate_resolutions.is_empty() {
+            warn!("Immediate res: {:?}", &immediate_resolutions);
         }
-        Ok(None)
+        for resolution in immediate_resolutions {
+            self.wfm
+                .notify_of_local_result(LocalResolution::LocalActivity(resolution))?;
+        }
+        Ok(())
     }
 
-    /// Returns `true` if autocompletion should be issued, which will actually cause us to end up
-    /// in [completion] again, at which point we'll start a new heartbeat timeout, which will
-    /// immediately trigger and thus finish the completion, forcing a new task as it should.
-    fn heartbeat_timeout(&mut self) -> bool {
-        if let Some(ref mut wait_dat) = self.waiting_on_la {
-            // Cancel the heartbeat timeout
-            wait_dat.heartbeat_timeout_task.abort();
-            if let Some((completion_dat, resp_chan)) = wait_dat.completion_dat.take() {
-                let compl = self.prepare_complete_resp(resp_chan, completion_dat, true);
-                // Immediately fulfill the completion since the run update will already have
-                // been replied to
-                compl.fulfill();
-            } else {
-                // Auto-reply WFT complete
-                return true;
-            }
-        } else {
-            // If a heartbeat timeout happened, we should always have been waiting on LAs
-            dbg_panic!("WFT heartbeat timeout fired but we were not waiting on any LAs");
+    fn reply_to_complete(
+        &self,
+        outcome: ActivationCompleteOutcome,
+        chan: Option<oneshot::Sender<ActivationCompleteResult>>,
+    ) {
+        if let Some(chan) = chan {
+            chan.send(ActivationCompleteResult {
+                most_recently_processed_event: self.most_recently_processed_event_number() as usize,
+                outcome,
+            })
+            .expect("Rcv half of activation reply not dropped");
         }
-        false
     }
 
-    fn send_update_response(&self, outcome: RunActionOutcome, no_wft: bool) {
-        let mut in_response_to_wft = false;
-        let (outgoing_activation, fulfillable_complete) = match outcome {
-            RunActionOutcome::AfterNewWFT(a) => {
-                in_response_to_wft = true;
-                (a, None)
-            }
-            RunActionOutcome::AfterCheckWork(a) => (a, None),
-            RunActionOutcome::AfterLocalResolution(f) => (None, f),
-            RunActionOutcome::AfterCompletion(f) => (None, f),
-            RunActionOutcome::AfterHeartbeatTimeout(a) => (a, None),
-        };
-        let mut more_pending_work = self.wfm.machines.has_pending_jobs();
-        // We don't want to consider there to be more local-only work to be done if there is no
-        // workflow task associated with the run right now. This can happen if, ex, we complete
-        // a local activity while waiting for server to send us the next WFT. Activating lang would
-        // be harmful at this stage, as there might be work returned in that next WFT which should
-        // be part of the next activation.
-        if no_wft {
-            more_pending_work = false;
-        }
-        self.update_tx
-            .send(RunUpdateResponse {
-                kind: RunUpdateResponseKind::Good(GoodRunUpdate {
-                    run_id: self.wfm.machines.run_id.clone(),
-                    outgoing_activation,
-                    fulfillable_complete,
-                    have_seen_terminal_event: self.wfm.machines.have_seen_terminal_event,
-                    more_pending_work,
-                    most_recently_processed_event_number: self.wfm.machines.last_processed_event
-                        as usize,
-                    in_response_to_wft,
-                }),
-                span: Span::current(),
-            })
-            .expect("Machine can send update");
+    /// Returns true if the handle is currently processing a WFT which contains a legacy query.
+    fn pending_work_is_legacy_query(&self) -> bool {
+        // Either we know because there is a pending legacy query, or it's already been drained and
+        // sent as an activation.
+        matches!(self.activation, Some(OutstandingActivation::LegacyQuery))
+            || self
+                .wft
+                .as_ref()
+                .map(|t| t.has_pending_legacy_query())
+                .unwrap_or_default()
+    }
+
+    fn most_recently_processed_event_number(&self) -> i64 {
+        self.wfm.machines.last_processed_event
+    }
+
+    fn activation_has_eviction(&mut self) -> bool {
+        self.activation
+            .map(OutstandingActivation::has_eviction)
+            .unwrap_or_default()
+    }
+
+    fn activation_has_only_eviction(&mut self) -> bool {
+        self.activation
+            .map(OutstandingActivation::has_only_eviction)
+            .unwrap_or_default()
+    }
+
+    fn run_id(&self) -> &str {
+        &self.wfm.machines.run_id
     }
 }
 
-fn start_heartbeat_timeout_task(
-    chan: UnboundedSender<Span>,
+/// Drains pending queries from the workflow task and appends them to the activation's jobs
+fn put_queries_in_act(act: &mut WorkflowActivation, wft: &mut OutstandingTask) {
+    // Nothing to do if there are no pending queries
+    if wft.pending_queries.is_empty() {
+        return;
+    }
+
+    let has_legacy = wft.has_pending_legacy_query();
+    // Cannot dispatch legacy query if there are any other jobs - which can happen if, ex, a local
+    // activity resolves while we've gotten a legacy query after heartbeating.
+    if has_legacy && !act.jobs.is_empty() {
+        return;
+    }
+
+    debug!(queries=?wft.pending_queries, "Dispatching queries");
+    let query_jobs = wft
+        .pending_queries
+        .drain(..)
+        .map(|q| workflow_activation_job::Variant::QueryWorkflow(q).into());
+    act.jobs.extend(query_jobs);
+}
+fn sink_heartbeat_timeout_start(
+    run_id: String,
+    sink: &dyn LocalActivityRequestSink,
     wft_start_time: Instant,
     wft_timeout: Duration,
-) -> JoinHandle<()> {
+) -> AbortHandle {
     // The heartbeat deadline is 80% of the WFT timeout
-    let wft_heartbeat_deadline =
-        wft_start_time.add(wft_timeout.mul_f32(WFT_HEARTBEAT_TIMEOUT_FRACTION));
-    task::spawn(async move {
-        tokio::time::sleep_until(wft_heartbeat_deadline.into()).await;
-        let _ = chan.send(Span::current());
-    })
-}
-
-enum RunActionOutcome {
-    AfterNewWFT(Option<ActivationOrAuto>),
-    AfterCheckWork(Option<ActivationOrAuto>),
-    AfterLocalResolution(Option<FulfillableActivationComplete>),
-    AfterCompletion(Option<FulfillableActivationComplete>),
-    AfterHeartbeatTimeout(Option<ActivationOrAuto>),
+    let deadline = wft_start_time.add(wft_timeout.mul_f32(WFT_HEARTBEAT_TIMEOUT_FRACTION));
+    let (abort_handle, abort_reg) = AbortHandle::new_pair();
+    sink.sink_reqs(vec![LocalActRequest::StartHeartbeatTimeout {
+        send_on_elapse: HeartbeatTimeoutMsg {
+            run_id,
+            span: Span::current(),
+        },
+        deadline,
+        abort_reg,
+    }]);
+    abort_handle
 }
 
-#[derive(derive_more::DebugCustom)]
-#[debug(fmt = "RunUpdateErr({:?})", source)]
-struct RunUpdateErr {
-    source: WFMachinesError,
-    complete_resp_chan: Option<oneshot::Sender<ActivationCompleteResult>>,
+/// If an activation completion needed to wait on LA completions (or heartbeat timeout) we use
+/// this struct to store the data we need to finish the completion once that has happened
+struct WaitingOnLAs {
+    wft_timeout: Duration,
+    /// If set, we are waiting for LAs to complete as part of a just-finished workflow activation.
+    /// If unset, we already had a heartbeat timeout and got a new WFT without any new work while
+    /// there are still incomplete LAs.
+    completion_dat: Option<(
+        CompletionDataForWFT,
+        Option<oneshot::Sender<ActivationCompleteResult>>,
+    )>,
+    /// Can be used to abort heartbeat timeouts
+    hb_timeout_handle: AbortHandle,
 }
-
-impl From<WFMachinesError> for RunUpdateErr {
-    fn from(e: WFMachinesError) -> Self {
-        RunUpdateErr {
-            source: e,
-            complete_resp_chan: None,
-        }
-    }
+#[derive(Debug)]
+struct CompletionDataForWFT {
+    task_token: TaskToken,
+    query_responses: Vec<QueryResult>,
+    has_pending_query: bool,
+    activation_was_only_eviction: bool,
 }
 
 /// Manages an instance of a [WorkflowMachines], which is not thread-safe, as well as other data
 /// associated with that specific workflow run.
-pub(crate) struct WorkflowManager {
+struct WorkflowManager {
     machines: WorkflowMachines,
     /// Is always `Some` in normal operation. Optional to allow for unit testing with the test
     /// workflow driver, which does not need to complete activations the normal way.
@@ -535,7 +1139,7 @@ pub(crate) struct WorkflowManager {
 impl WorkflowManager {
     /// Create a new workflow manager given workflow history and execution info as would be found
     /// in [PollWorkflowTaskQueueResponse]
-    pub fn new(
+    fn new(
         history: HistoryUpdate,
         namespace: String,
         workflow_id: String,
@@ -560,7 +1164,7 @@ impl WorkflowManager {
     }
 
     #[cfg(test)]
-    pub const fn new_from_machines(workflow_machines: WorkflowMachines) -> Self {
+    const fn new_from_machines(workflow_machines: WorkflowMachines) -> Self {
         Self {
             machines: workflow_machines,
             command_sink: None,
@@ -571,12 +1175,15 @@ impl WorkflowManager {
     ///
     /// Should only be called when a workflow has caught up on replay (or is just beginning). It
     /// will return a workflow activation if one is needed.
-    async fn feed_history_from_server(
-        &mut self,
-        update: HistoryUpdate,
-    ) -> Result<WorkflowActivation> {
-        self.machines.new_history_from_server(update).await?;
-        self.get_next_activation().await
+    fn feed_history_from_server(&mut self, update: HistoryUpdate) -> Result<WorkflowActivation> {
+        self.machines.new_history_from_server(update)?;
+        self.get_next_activation()
+    }
+
+    /// Update the machines with some events from fetching another page of history. Does *not*
+    /// attempt to pull the next activation, unlike [Self::feed_history_from_server].
+    fn feed_history_from_new_page(&mut self, update: HistoryUpdate) -> Result<()> {
+        self.machines.new_history_from_server(update)
     }
 
     /// Let this workflow know that something we've been waiting locally on has resolved, like a
@@ -593,27 +1200,33 @@ impl WorkflowManager {
     ///
     /// Callers may also need to call [get_server_commands] after this to issue any pending commands
     /// to the server.
-    async fn get_next_activation(&mut self) -> Result<WorkflowActivation> {
+    fn get_next_activation(&mut self) -> Result<WorkflowActivation> {
         // First check if there are already some pending jobs, which can be a result of replay.
         let activation = self.machines.get_wf_activation();
         if !activation.jobs.is_empty() {
             return Ok(activation);
         }
 
-        self.machines.apply_next_wft_from_history().await?;
+        self.machines.apply_next_wft_from_history()?;
         Ok(self.machines.get_wf_activation())
     }
 
+    /// Returns true if machines are ready to apply the next WFT sequence, false if events will need
+    /// to be fetched in order to create a complete update with the entire next WFT sequence.
+    pub(crate) fn ready_to_apply_next_wft(&self) -> bool {
+        self.machines.ready_to_apply_next_wft()
+    }
+
     /// If there are no pending jobs for the workflow, apply the next workflow task and check
     /// again if there are any jobs. Importantly, does not *drain* jobs.
     ///
     /// Returns true if there are jobs (before or after applying the next WFT).
-    async fn apply_next_task_if_ready(&mut self) -> Result<bool> {
+    fn apply_next_task_if_ready(&mut self) -> Result<bool> {
         if self.machines.has_pending_jobs() {
             return Ok(true);
         }
         loop {
-            let consumed_events = self.machines.apply_next_wft_from_history().await?;
+            let consumed_events = self.machines.apply_next_wft_from_history()?;
 
             if consumed_events == 0 || !self.machines.replaying || self.machines.has_pending_jobs()
             {
@@ -643,13 +1256,61 @@ impl WorkflowManager {
 
     /// Feed the workflow machines new commands issued by the executing workflow code, and iterate
     /// the machines.
-    async fn push_commands(&mut self, cmds: Vec<WFCommand>) -> Result<()> {
+    fn push_commands_and_iterate(&mut self, cmds: Vec<WFCommand>) -> Result<()> {
         if let Some(cs) = self.command_sink.as_mut() {
             cs.send(cmds).map_err(|_| {
                 WFMachinesError::Fatal("Internal error buffering workflow commands".to_string())
             })?;
         }
-        self.machines.iterate_machines().await?;
+        self.machines.iterate_machines()?;
         Ok(())
     }
 }
+
+#[derive(Debug)]
+struct FulfillableActivationComplete {
+    result: ActivationCompleteResult,
+    resp_chan: Option<oneshot::Sender<ActivationCompleteResult>>,
+}
+impl FulfillableActivationComplete {
+    fn fulfill(self) {
+        if let Some(resp_chan) = self.resp_chan {
+            let _ = resp_chan.send(self.result);
+        }
+    }
+}
+
+#[derive(Debug)]
+struct RunActivationCompletion {
+    task_token: TaskToken,
+    start_time: Instant,
+    commands: Vec<WFCommand>,
+    activation_was_eviction: bool,
+    activation_was_only_eviction: bool,
+    has_pending_query: bool,
+    query_responses: Vec<QueryResult>,
+    /// Used to notify the worker when the completion is done processing and the completion can
+    /// unblock. Must always be `Some` when initialized.
+    resp_chan: Option<oneshot::Sender<ActivationCompleteResult>>,
+}
+#[derive(Debug, derive_more::From)]
+enum ActOrFulfill {
+    OutgoingAct(Option<ActivationOrAuto>),
+    FulfillableComplete(Option<FulfillableActivationComplete>),
+}
+
+#[derive(derive_more::DebugCustom)]
+#[debug(fmt = "RunUpdateErr({source:?})")]
+struct RunUpdateErr {
+    source: WFMachinesError,
+    complete_resp_chan: Option<oneshot::Sender<ActivationCompleteResult>>,
+}
+
+impl From<WFMachinesError> for RunUpdateErr {
+    fn from(e: WFMachinesError) -> Self {
+        RunUpdateErr {
+            source: e,
+            complete_resp_chan: None,
+        }
+    }
+}