WIP(not working)

Author: LiaCastaneda

Cargo.lock

@@ -55,6 +55,7 @@ kube = { version = "1.1", features = ["derive", "runtime"] }
 log = "0.4"
 rand = "0.8"
 uuid = { version = "1.6", features = ["v4"] }
+tower = "0.5"
 
 object_store = { version = "0.12.0", features = [
   "aws",

Cargo.toml

@@ -2,7 +2,7 @@ use anyhow::Result;
 use clap::Parser;
 use distributed_datafusion::{
     friendly::new_friendly_name, proxy_service::DDProxyService, setup,
-    worker_service::DDWorkerService,
+    worker_service::DDWorkerService, worker_discovery::{EnvDiscovery, WorkerDiscovery},
 };
 
 #[derive(Parser)]
@@ -30,10 +30,16 @@ async fn main() -> Result<()> {
     setup();
 
     let args = Args::parse();
-
+    
     match args.mode.as_str() {
         "proxy" => {
-            let service = DDProxyService::new(new_friendly_name()?, args.port, None).await?;
+            // TODO: put the k8s or ENV decision behind some flag, WARNING: this will kick the discovery so workers should be up
+            let discovery: Arc<dyn WorkerDiscovery> = if std::env::var("DD_WORKER_ADDRESSES").is_ok() {
+                Arc::new(EnvDiscovery::new().await?)
+            } else {
+                Arc::new(K8sDiscovery::new().await?)
+            };
+            let service = DDProxyService::new(new_friendly_name()?, args.port,discovery, None).await?;
             service.serve().await?;
         }
         "worker" => {

src/bin/distributed-datafusion.rs

@@ -43,6 +43,9 @@ pub mod util;
 pub mod vocab;
 pub mod worker_discovery;
 pub mod worker_service;
+pub mod transport_traits;
+pub mod transport;
+pub mod test_worker;
 
 #[cfg(not(target_env = "msvc"))]
 use tikv_jemallocator::Jemalloc;

src/lib.rs

@@ -1,7 +1,5 @@
 use std::{
-    collections::HashMap,
-    env,
-    sync::{Arc, LazyLock},
+    collections::HashMap, env, sync::{Arc, LazyLock}
 };
 
 use anyhow::{anyhow, Context};
@@ -30,19 +28,10 @@ use itertools::Itertools;
 use prost::Message;
 
 use crate::{
-    analyze::{DistributedAnalyzeExec, DistributedAnalyzeRootExec},
-    isolator::PartitionIsolatorExec,
-    logging::{debug, error, info, trace},
-    max_rows::MaxRowsExec,
-    physical::DDStageOptimizerRule,
-    result::{DDError, Result},
-    stage::DDStageExec,
-    stage_reader::{DDStageReaderExec, QueryId},
-    util::{display_plan_with_partition_counts, get_client, physical_plan_to_bytes, wait_for},
-    vocab::{
+    analyze::{DistributedAnalyzeExec, DistributedAnalyzeRootExec}, isolator::PartitionIsolatorExec, logging::{debug, info, trace, error}, max_rows::MaxRowsExec, physical::DDStageOptimizerRule, result::{DDError, Result}, stage::DDStageExec, stage_reader::{DDStageReaderExec, QueryId}, transport::WorkerTransport, util::{display_plan_with_partition_counts, physical_plan_to_bytes, wait_for}, vocab::{
         Addrs, CtxAnnotatedOutputs, CtxHost, CtxPartitionGroup, CtxStageAddrs, CtxStageId, DDTask,
         Host, Hosts, PartitionAddrs, StageAddrs,
-    },
+    }
 };
 
 #[derive(Debug)]
@@ -439,109 +428,84 @@ pub fn add_distributed_analyze(
 pub async fn distribute_stages(
     query_id: &str,
     stages: Vec<DDStage>,
-    worker_addrs: Vec<Host>,
+    workers: Vec<(Host, Arc<dyn WorkerTransport>)>,
     codec: &dyn PhysicalExtensionCodec,
 ) -> Result<(Addrs, Vec<DDTask>)> {
-    // map of worker name to address
-    // FIXME: use types over tuples of strings, as we can accidently swap them and
-    // not know
-
-    // a map of worker name to host
-    let mut workers: HashMap<String, Host> = worker_addrs
-        .iter()
-        .map(|host| (host.name.clone(), host.clone()))
+    // materialise a name-keyed map so we can remove “bad” workers on each retry
+    let mut valid_workers: HashMap<_, _> = workers
+        .into_iter()
+        .map(|(h, tx)| (h.name.clone(), (h, tx)))
         .collect();
 
     for attempt in 0..3 {
-        if workers.is_empty() {
+        if valid_workers.is_empty() {
             return Err(anyhow!("No workers available to distribute stages").into());
         }
 
-        // all stages to workers
-        let (task_datas, final_addrs) =
-            assign_to_workers(query_id, &stages, workers.values().collect(), codec)?;
+        let current: Vec<_> = valid_workers.values().cloned().collect();
+        let (tasks, final_addrs, tx_host_pairs) =
+            assign_to_workers(query_id, &stages, current, codec)?;
+
+        match try_distribute_tasks(&tasks, &tx_host_pairs).await {
+            Ok(_) => return Ok((final_addrs, tasks)),
 
-        // we retry this a few times to ensure that the workers are ready
-        // and can accept the stages
-        match try_distribute_tasks(&task_datas).await {
-            Ok(_) => return Ok((final_addrs, task_datas)),
-            Err(DDError::WorkerCommunicationError(bad_worker)) => {
+            // remove the poisoned worker and retry on the non poisoned workers
+            Err(DDError::WorkerCommunicationError(bad_host)) => {
                 error!(
-                    "distribute stages for query {query_id} attempt {attempt} failed removing \
-                     worker {bad_worker}. Retrying..."
+                    "distribute_stages: attempt {attempt} – \
+                     worker {} failed; will retry without it",
+                    bad_host.name
                 );
-                // if we cannot communicate with a worker, we remove it from the list of workers
-                workers.remove(&bad_worker.name);
+                valid_workers.remove(&bad_host.name);
             }
+
+            // any other error is terminal
             Err(e) => return Err(e),
         }
-        if attempt == 2 {
-            return Err(
-                anyhow!("Failed to distribute query {query_id} stages after 3 attempts").into(),
-            );
-        }
     }
-    unreachable!()
+
+    unreachable!("retry loop exits on success or early return on error");
 }
 
 /// try to distribute the stages to the workers, if we cannot communicate with a
 /// worker return it as the element in the Err
-async fn try_distribute_tasks(task_datas: &[DDTask]) -> Result<()> {
-    // we can use the stage data to distribute the stages to workers
-    for task_data in task_datas {
+async fn try_distribute_tasks(
+    tasks: &[DDTask],
+    tx_host_pairs: &[(Arc<dyn WorkerTransport>, Host)],
+) -> Result<()> {
+    for ((task, (tx, host))) in tasks.iter().zip(tx_host_pairs) {
         trace!(
-            "Distributing Task: stage_id {}, pg: {:?} to worker: {:?}",
-            task_data.stage_id,
-            task_data.partition_group,
-            task_data.assigned_host
+            "Sending stage {} / pg {:?} to {}",
+            task.stage_id,
+            task.partition_group,
+            host
         );
 
-        // populate its child stages
-        let mut stage_data = task_data.clone();
-        stage_data.stage_addrs = Some(get_stage_addrs_from_tasks(
-            &stage_data.child_stage_ids,
-            task_datas,
+        // embed the StageAddrs of all children before shipping
+        let mut stage = task.clone();
+        stage.stage_addrs = Some(get_stage_addrs_from_tasks(
+            &stage.child_stage_ids,
+            tasks,
         )?);
 
-        let host = stage_data
-            .assigned_host
-            .clone()
-            .context("Assigned host is missing for task data")?;
-
-        let mut client = match get_client(&host) {
-            Ok(client) => client,
-            Err(e) => {
-                error!("Couldn't not communicate with worker {e:#?}");
-                return Err(DDError::WorkerCommunicationError(
-                    host.clone(), // here
-                ));
-            }
-        };
-
-        let mut buf = vec![];
-        stage_data
-            .encode(&mut buf)
-            .context("Failed to encode stage data to buf")?;
+        let mut buf = Vec::new();
+        stage.encode(&mut buf).map_err(anyhow::Error::from)?;
 
         let action = Action {
-            r#type: "add_plan".to_string(),
+            r#type: "add_plan".into(),
             body: buf.into(),
         };
 
-        let mut response = client
+        // gRPC call, if it fails, transport poisons itself on failure and removes the address from the registry
+        let mut stream = tx
             .do_action(action)
             .await
-            .context("Failed to send action to worker")?;
+            .map_err(|_| DDError::WorkerCommunicationError(host.clone()))?;
 
-        // consume this empty response to ensure the action was successful
-        while let Some(_res) = response
-            .try_next()
-            .await
-            .context("error consuming do_action response")?
-        {
-            // we don't care about the response, just that it was successful
-        }
-        trace!("do action success for stage_id: {}", stage_data.stage_id);
+        // drain the (empty) response – ensures the worker actually accepted it
+        while stream.try_next().await? != None {}
+
+        trace!("stage {} delivered to {}", stage.stage_id, host);
     }
     Ok(())
 }
@@ -552,40 +516,35 @@ async fn try_distribute_tasks(task_datas: &[DDTask]) -> Result<()> {
 fn assign_to_workers(
     query_id: &str,
     stages: &[DDStage],
-    worker_addrs: Vec<&Host>,
+    workers: Vec<(Host, Arc<dyn WorkerTransport>)>,
     codec: &dyn PhysicalExtensionCodec,
-) -> Result<(Vec<DDTask>, Addrs)> {
-    let mut task_datas = vec![];
-    let mut worker_idx = 0;
+) -> Result<(Vec<DDTask>, Addrs, Vec<(Arc<dyn WorkerTransport>, Host)>)> {
+    let mut task_datas   = Vec::new();
+    let mut tx_host_pairs = Vec::new();
 
-    trace!(
-        "assigning stages: {:?}",
-        stages
-            .iter()
-            .map(|s| format!("stage_id: {}, pgs:{:?}", s.stage_id, s.partition_groups))
-            .join(",\n")
-    );
+    // round-robin scheduler
+    let mut idx = 0;
+    let n_workers = workers.len();
 
-    // keep track of which worker has the root of the plan tree (highest stage
-    // number)
-    let mut max_stage_id = -1;
+    // keep track of where the root of the plan will live (highest stage id)
+    let mut max_stage_id: i64 = -1;
     let mut final_addrs = Addrs::default();
 
     for stage in stages {
-        for partition_group in stage.partition_groups.iter() {
+        for pg in &stage.partition_groups {
             let plan_bytes = physical_plan_to_bytes(stage.plan.clone(), codec)?;
 
-            let host = worker_addrs[worker_idx].clone();
-            worker_idx = (worker_idx + 1) % worker_addrs.len();
+            // pick next worker
+            let (host, tx) = workers[idx].clone();
+            idx = (idx + 1) % n_workers;
 
-            if stage.stage_id as isize > max_stage_id {
-                // this wasn't the last stage
-                max_stage_id = stage.stage_id as isize;
+            // remember which host serves the final stage
+            if stage.stage_id as i64 > max_stage_id {
+                max_stage_id = stage.stage_id as i64;
                 final_addrs.clear();
             }
-            if stage.stage_id as isize == max_stage_id {
-                for part in partition_group.iter() {
-                    // we are the final stage, so we will be the one to serve this partition
+            if stage.stage_id as i64 == max_stage_id {
+                for part in pg {
                     final_addrs
                         .entry(stage.stage_id)
                         .or_default()
@@ -595,22 +554,24 @@ fn assign_to_workers(
                 }
             }
 
-            let task_data = DDTask {
-                query_id: query_id.to_string(),
+            task_datas.push(DDTask {
+                query_id: query_id.to_owned(),
                 stage_id: stage.stage_id,
                 plan_bytes,
-                partition_group: partition_group.to_vec(),
-                child_stage_ids: stage.child_stage_ids().unwrap_or_default().to_vec(),
-                stage_addrs: None, // will be calculated and filled in later
+                partition_group: pg.clone(),
+                child_stage_ids: stage.child_stage_ids().unwrap_or_default(),
+                stage_addrs: None, // filled in later
                 num_output_partitions: stage.plan.output_partitioning().partition_count() as u64,
                 full_partitions: stage.full_partitions,
-                assigned_host: Some(host),
-            };
-            task_datas.push(task_data);
+                assigned_host: Some(host.clone()),
+            });
+
+            // keep the order **exactly** aligned with task_datas
+            tx_host_pairs.push((tx, host));
         }
     }
 
-    Ok((task_datas, final_addrs))
+    Ok((task_datas, final_addrs, tx_host_pairs))
 }
 
 fn get_stage_addrs_from_tasks(target_stage_ids: &[u64], stages: &[DDTask]) -> Result<StageAddrs> {

src/planning.rs

@@ -50,31 +50,30 @@ use crate::{
     stage_reader::DDStageReaderExec,
     util::{display_plan_with_partition_counts, get_addrs, start_up},
     vocab::{Addrs, Host},
-    worker_discovery::get_worker_addresses,
+    worker_discovery::{ WorkerDiscovery},
 };
 
 pub struct DDProxyHandler {
     /// our host info, useful for logging
     pub host: Host,
 
     pub planner: QueryPlanner,
+    pub discovery: Arc<dyn WorkerDiscovery>,
 
     /// Optional customizer for our context and proto serde
     pub customizer: Option<Arc<dyn Customizer>>,
 }
 
 impl DDProxyHandler {
-    pub fn new(name: String, addr: String, customizer: Option<Arc<dyn Customizer>>) -> Self {
-        // call this function to bootstrap the worker discovery mechanism
-        get_worker_addresses().expect("Could not get worker addresses upon startup");
-
+    pub fn new(name: String, addr: String, discovery: Arc<dyn WorkerDiscovery>, customizer: Option<Arc<dyn Customizer>>) -> Self {
         let host = Host {
             name: name.clone(),
             addr: addr.clone(),
         };
         Self {
             host: host.clone(),
-            planner: QueryPlanner::new(customizer.clone()),
+            planner: QueryPlanner::new(customizer.clone(), discovery.clone()),
+            discovery,
             customizer,
         }
     }
@@ -292,6 +291,7 @@ impl DDProxyService {
     pub async fn new(
         name: String,
         port: usize,
+        discovery: Arc<dyn WorkerDiscovery>,
         ctx_customizer: Option<Arc<dyn Customizer>>,
     ) -> Result<Self> {
         debug!("Creating DDProxyService!");
@@ -305,7 +305,7 @@ impl DDProxyService {
 
         info!("DDProxyService bound to {addr}");
 
-        let handler = Arc::new(DDProxyHandler::new(name, addr.clone(), ctx_customizer));
+        let handler = Arc::new(DDProxyHandler::new(name, addr.clone(), discovery,  ctx_customizer));
 
         Ok(Self {
             listener,