+// Copyright (C) The Arvados Authors. All rights reserved.
+//
+// SPDX-License-Identifier: AGPL-3.0
+
package main
+// Dispatcher service for Crunch that runs containers locally.
+
import (
+ "context"
"flag"
- "git.curoverse.com/arvados.git/sdk/go/arvadosclient"
"log"
"os"
"os/exec"
"sync"
"syscall"
"time"
+
+ "git.curoverse.com/arvados.git/sdk/go/arvados"
+ "git.curoverse.com/arvados.git/sdk/go/arvadosclient"
+ "git.curoverse.com/arvados.git/sdk/go/dispatch"
)
func main() {
}
var (
- arv arvadosclient.ArvadosClient
runningCmds map[string]*exec.Cmd
runningCmdsMutex sync.Mutex
waitGroup sync.WaitGroup
- doneProcessing chan bool
- sigChan chan os.Signal
+ crunchRunCommand *string
)
func doMain() error {
10,
"Interval in seconds to poll for queued containers")
- priorityPollInterval := flags.Int(
- "container-priority-poll-interval",
- 60,
- "Interval in seconds to check priority of a dispatched container")
-
- crunchRunCommand := flags.String(
+ crunchRunCommand = flags.String(
"crunch-run-command",
"/usr/bin/crunch-run",
"Crunch command to run container")
// Parse args; omit the first arg which is the command name
flags.Parse(os.Args[1:])
- var err error
- arv, err = arvadosclient.MakeArvadosClient()
+ runningCmds = make(map[string]*exec.Cmd)
+
+ arv, err := arvadosclient.MakeArvadosClient()
if err != nil {
+ log.Printf("Error making Arvados client: %v", err)
return err
}
+ arv.Retries = 25
- // Channel to terminate
- doneProcessing = make(chan bool)
+ dispatcher := dispatch.Dispatcher{
+ Arv: arv,
+ RunContainer: run,
+ PollPeriod: time.Duration(*pollInterval) * time.Second,
+ }
- // Map of running crunch jobs
- runningCmds = make(map[string]*exec.Cmd)
+ ctx, cancel := context.WithCancel(context.Background())
+ err = dispatcher.Run(ctx)
+ if err != nil {
+ return err
+ }
- // Graceful shutdown
- sigChan = make(chan os.Signal, 1)
- signal.Notify(sigChan, syscall.SIGINT, syscall.SIGTERM, syscall.SIGQUIT)
- go func(sig <-chan os.Signal) {
- for sig := range sig {
- log.Printf("Caught signal: %v", sig)
- doneProcessing <- true
- }
- }(sigChan)
+ c := make(chan os.Signal, 1)
+ signal.Notify(c, os.Interrupt, syscall.SIGTERM, syscall.SIGQUIT)
+ sig := <-c
+ log.Printf("Received %s, shutting down", sig)
+ signal.Stop(c)
- // Run all queued containers
- runQueuedContainers(*pollInterval, *priorityPollInterval, *crunchRunCommand)
+ cancel()
+ runningCmdsMutex.Lock()
// Finished dispatching; interrupt any crunch jobs that are still running
- for uuid, cmd := range runningCmds {
- go func(uuid string) {
- cmd.Process.Signal(os.Interrupt)
- if _, err := cmd.Process.Wait(); err != nil {
- log.Printf("Error while waiting for crunch job to finish for %v: %q", uuid, err)
- }
- }(uuid)
+ for _, cmd := range runningCmds {
+ cmd.Process.Signal(os.Interrupt)
}
+ runningCmdsMutex.Unlock()
// Wait for all running crunch jobs to complete / terminate
waitGroup.Wait()
return nil
}
-// Poll for queued containers using pollInterval.
-// Invoke dispatchLocal for each ticker cycle, which will run all the queued containers.
-//
-// Any errors encountered are logged but the program would continue to run (not exit).
-// This is because, once one or more crunch jobs are running,
-// we would need to wait for them complete.
-func runQueuedContainers(pollInterval, priorityPollInterval int, crunchRunCommand string) {
- ticker := time.NewTicker(time.Duration(pollInterval) * time.Second)
-
- for {
- select {
- case <-ticker.C:
- dispatchLocal(priorityPollInterval, crunchRunCommand)
- case <-doneProcessing:
- ticker.Stop()
- return
- }
- }
-}
-
-// Container data
-type Container struct {
- UUID string `json:"uuid"`
- State string `json:"state"`
- Priority int `json:"priority"`
+func startFunc(container arvados.Container, cmd *exec.Cmd) error {
+ return cmd.Start()
}
-// ContainerList is a list of the containers from api
-type ContainerList struct {
- Items []Container `json:"items"`
-}
+var startCmd = startFunc
-// Get the list of queued containers from API server and invoke run for each container.
-func dispatchLocal(priorityPollInterval int, crunchRunCommand string) {
- params := arvadosclient.Dict{
- "filters": [][]string{[]string{"state", "=", "Queued"}},
- }
-
- var containers ContainerList
- err := arv.List("containers", params, &containers)
- if err != nil {
- log.Printf("Error getting list of queued containers: %q", err)
- return
- }
-
- for i := 0; i < len(containers.Items); i++ {
- log.Printf("About to run queued container %v", containers.Items[i].UUID)
- // Run the container
- go run(containers.Items[i].UUID, crunchRunCommand, priorityPollInterval)
- }
-}
-
-// Run queued container:
-// Set container state to locked (TBD)
-// Run container using the given crunch-run command
-// Set the container state to Running
-// If the container priority becomes zero while crunch job is still running, terminate it.
-func run(uuid string, crunchRunCommand string, priorityPollInterval int) {
- cmd := exec.Command(crunchRunCommand, uuid)
-
- cmd.Stdin = nil
- cmd.Stderr = os.Stderr
- cmd.Stdout = os.Stderr
- if err := cmd.Start(); err != nil {
- log.Printf("Error running container for %v: %q", uuid, err)
- return
- }
-
- // Add this crunch job to the list of runningCmds
- runningCmdsMutex.Lock()
- runningCmds[uuid] = cmd
- runningCmdsMutex.Unlock()
-
- log.Printf("Started container run for %v", uuid)
-
- // Add this crunch job to waitGroup
- waitGroup.Add(1)
- defer waitGroup.Done()
-
- // Update container status to Running
- err := arv.Update("containers", uuid,
- arvadosclient.Dict{
- "container": arvadosclient.Dict{"state": "Running"}},
- nil)
- if err != nil {
- log.Printf("Error updating container state to 'Running' for %v: %q", uuid, err)
- }
-
- // A goroutine to terminate the runner if container priority becomes zero
- priorityTicker := time.NewTicker(time.Duration(priorityPollInterval) * time.Second)
- go func() {
- for {
- select {
- case <-priorityTicker.C:
- var container Container
- err := arv.Get("containers", uuid, nil, &container)
- if err != nil {
- log.Printf("Error getting container info for %v: %q", uuid, err)
- } else {
- if container.Priority == 0 {
- priorityTicker.Stop()
+// Run a container.
+//
+// If the container is Locked, start a new crunch-run process and wait until
+// crunch-run completes. If the priority is set to zero, set an interrupt
+// signal to the crunch-run process.
+//
+// If the container is in any other state, or is not Complete/Cancelled after
+// crunch-run terminates, mark the container as Cancelled.
+func run(dispatcher *dispatch.Dispatcher,
+ container arvados.Container,
+ status <-chan arvados.Container) {
+
+ uuid := container.UUID
+
+ if container.State == dispatch.Locked {
+ waitGroup.Add(1)
+
+ cmd := exec.Command(*crunchRunCommand, uuid)
+ cmd.Stdin = nil
+ cmd.Stderr = os.Stderr
+ cmd.Stdout = os.Stderr
+
+ log.Printf("Starting container %v", uuid)
+
+ // Add this crunch job to the list of runningCmds only if we
+ // succeed in starting crunch-run.
+
+ runningCmdsMutex.Lock()
+ if err := startCmd(container, cmd); err != nil {
+ runningCmdsMutex.Unlock()
+ log.Printf("Error starting %v for %v: %q", *crunchRunCommand, uuid, err)
+ dispatcher.UpdateState(uuid, dispatch.Cancelled)
+ } else {
+ runningCmds[uuid] = cmd
+ runningCmdsMutex.Unlock()
+
+ // Need to wait for crunch-run to exit
+ done := make(chan struct{})
+
+ go func() {
+ if _, err := cmd.Process.Wait(); err != nil {
+ log.Printf("Error while waiting for crunch job to finish for %v: %q", uuid, err)
+ }
+ log.Printf("sending done")
+ done <- struct{}{}
+ }()
+
+ Loop:
+ for {
+ select {
+ case <-done:
+ break Loop
+ case c := <-status:
+ // Interrupt the child process if priority changes to 0
+ if (c.State == dispatch.Locked || c.State == dispatch.Running) && c.Priority == 0 {
+ log.Printf("Sending SIGINT to pid %d to cancel container %v", cmd.Process.Pid, uuid)
cmd.Process.Signal(os.Interrupt)
- runningCmdsMutex.Lock()
- delete(runningCmds, uuid)
- runningCmdsMutex.Unlock()
- return
}
}
}
- }
- }()
+ close(done)
- // Wait for the crunch job to exit
- if _, err := cmd.Process.Wait(); err != nil {
- log.Printf("Error while waiting for crunch job to finish for %v: %q", uuid, err)
+ log.Printf("Finished container run for %v", uuid)
+
+ // Remove the crunch job from runningCmds
+ runningCmdsMutex.Lock()
+ delete(runningCmds, uuid)
+ runningCmdsMutex.Unlock()
+ }
+ waitGroup.Done()
}
- // Remove the crunch job to runningCmds
- runningCmdsMutex.Lock()
- delete(runningCmds, uuid)
- runningCmdsMutex.Unlock()
+ // If the container is not finalized, then change it to "Cancelled".
+ err := dispatcher.Arv.Get("containers", uuid, nil, &container)
+ if err != nil {
+ log.Printf("Error getting final container state: %v", err)
+ }
+ if container.State == dispatch.Locked || container.State == dispatch.Running {
+ log.Printf("After %s process termination, container state for %v is %q. Updating it to %q",
+ *crunchRunCommand, container.State, uuid, dispatch.Cancelled)
+ dispatcher.UpdateState(uuid, dispatch.Cancelled)
+ }
- priorityTicker.Stop()
-
- // The container state should be 'Complete'
- var container Container
- err = arv.Get("containers", uuid, nil, &container)
- if container.State == "Running" {
- log.Printf("After crunch-run process termination, the state is still 'Running' for %v. Updating it to 'Complete'", uuid)
- err = arv.Update("containers", uuid,
- arvadosclient.Dict{
- "container": arvadosclient.Dict{"state": "Complete"}},
- nil)
- if err != nil {
- log.Printf("Error updating container state to Complete for %v: %q", uuid, err)
- }
+ // drain any subsequent status changes
+ for range status {
}
+
+ log.Printf("Finalized container %v", uuid)
}