X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/9333c9f65503d86c12776e0bc8bfcb6fc07dc79c..443a0b96316ed46600dc5035193adae6ac4d1f74:/sdk/go/dispatch/dispatch.go diff --git a/sdk/go/dispatch/dispatch.go b/sdk/go/dispatch/dispatch.go index 5341369d01..fdb52e510b 100644 --- a/sdk/go/dispatch/dispatch.go +++ b/sdk/go/dispatch/dispatch.go @@ -1,3 +1,7 @@ +// Copyright (C) The Arvados Authors. All rights reserved. +// +// SPDX-License-Identifier: Apache-2.0 + // Package dispatch is a helper library for building Arvados container // dispatchers. package dispatch @@ -5,12 +9,12 @@ package dispatch import ( "context" "fmt" - "log" "sync" "time" "git.curoverse.com/arvados.git/sdk/go/arvados" "git.curoverse.com/arvados.git/sdk/go/arvadosclient" + "github.com/sirupsen/logrus" ) const ( @@ -21,9 +25,21 @@ const ( Cancelled = arvados.ContainerStateCancelled ) +type Logger interface { + Printf(string, ...interface{}) + Warnf(string, ...interface{}) + Debugf(string, ...interface{}) +} + +// Dispatcher struct type Dispatcher struct { Arv *arvadosclient.ArvadosClient + Logger Logger + + // Batch size for container queries + BatchSize int64 + // Queue polling frequency PollPeriod time.Duration @@ -36,7 +52,7 @@ type Dispatcher struct { auth arvados.APIClientAuthorization mtx sync.Mutex - running map[string]*runTracker + trackers map[string]*runTracker throttle throttle } @@ -57,6 +73,10 @@ type DispatchFunc func(*Dispatcher, arvados.Container, <-chan arvados.Container) // dispatcher's token. When a new one appears, Run calls RunContainer // in a new goroutine. func (d *Dispatcher) Run(ctx context.Context) error { + if d.Logger == nil { + d.Logger = logrus.StandardLogger() + } + err := d.Arv.Call("GET", "api_client_authorizations", "", "current", nil, &d.auth) if err != nil { return fmt.Errorf("error getting my token UUID: %v", err) @@ -67,85 +87,154 @@ func (d *Dispatcher) Run(ctx context.Context) error { poll := time.NewTicker(d.PollPeriod) defer poll.Stop() + if d.BatchSize == 0 { + d.BatchSize = 100 + } + for { - d.checkForUpdates([][]interface{}{ - {"uuid", "in", d.runningUUIDs()}}) - d.checkForUpdates([][]interface{}{ - {"locked_by_uuid", "=", d.auth.UUID}, - {"uuid", "not in", d.runningUUIDs()}}) - d.checkForUpdates([][]interface{}{ - {"state", "=", Queued}, - {"priority", ">", "0"}, - {"uuid", "not in", d.runningUUIDs()}}) select { case <-poll.C: - continue + break case <-ctx.Done(): return ctx.Err() } - } -} -func (d *Dispatcher) runningUUIDs() []string { - d.mtx.Lock() - defer d.mtx.Unlock() - if len(d.running) == 0 { - // API bug: ["uuid", "not in", []] does not match everything - return []string{"X"} - } - uuids := make([]string, 0, len(d.running)) - for x := range d.running { - uuids = append(uuids, x) + todo := make(map[string]*runTracker) + d.mtx.Lock() + // Make a copy of trackers + for uuid, tracker := range d.trackers { + todo[uuid] = tracker + } + d.mtx.Unlock() + + // Containers I currently own (Locked/Running) + querySuccess := d.checkForUpdates([][]interface{}{ + {"locked_by_uuid", "=", d.auth.UUID}}, todo) + + // Containers I should try to dispatch + querySuccess = d.checkForUpdates([][]interface{}{ + {"state", "=", Queued}, + {"priority", ">", "0"}}, todo) && querySuccess + + if !querySuccess { + // There was an error in one of the previous queries, + // we probably didn't get updates for all the + // containers we should have. Don't check them + // individually because it may be expensive. + continue + } + + // Containers I know about but didn't fall into the + // above two categories (probably Complete/Cancelled) + var missed []string + for uuid := range todo { + missed = append(missed, uuid) + } + + for len(missed) > 0 { + var batch []string + if len(missed) > 20 { + batch = missed[0:20] + missed = missed[20:] + } else { + batch = missed + missed = missed[0:0] + } + querySuccess = d.checkForUpdates([][]interface{}{ + {"uuid", "in", batch}}, todo) && querySuccess + } + + if !querySuccess { + // There was an error in one of the previous queries, we probably + // didn't see all the containers we should have, so don't shut down + // the missed containers. + continue + } + + // Containers that I know about that didn't show up in any + // query should be let go. + for uuid, tracker := range todo { + d.Logger.Printf("Container %q not returned by any query, stopping tracking.", uuid) + tracker.close() + } + } - return uuids } // Start a runner in a new goroutine, and send the initial container // record to its updates channel. func (d *Dispatcher) start(c arvados.Container) *runTracker { - tracker := &runTracker{updates: make(chan arvados.Container, 1)} + tracker := &runTracker{ + updates: make(chan arvados.Container, 1), + logger: d.Logger, + } tracker.updates <- c go func() { d.RunContainer(d, c, tracker.updates) - + // RunContainer blocks for the lifetime of the container. When + // it returns, the tracker should delete itself. d.mtx.Lock() - delete(d.running, c.UUID) + delete(d.trackers, c.UUID) d.mtx.Unlock() }() return tracker } -func (d *Dispatcher) checkForUpdates(filters [][]interface{}) { +func (d *Dispatcher) checkForUpdates(filters [][]interface{}, todo map[string]*runTracker) bool { + var countList arvados.ContainerList params := arvadosclient.Dict{ "filters": filters, - "order": []string{"priority desc"}, - "limit": "1000"} - - var list arvados.ContainerList - err := d.Arv.List("containers", params, &list) + "count": "exact", + "limit": 0, + "order": []string{"priority desc"}} + err := d.Arv.List("containers", params, &countList) if err != nil { - log.Printf("Error getting list of containers: %q", err) - return + d.Logger.Warnf("error getting count of containers: %q", err) + return false } + itemsAvailable := countList.ItemsAvailable + params = arvadosclient.Dict{ + "filters": filters, + "count": "none", + "limit": d.BatchSize, + "order": []string{"priority desc"}} + offset := 0 + for { + params["offset"] = offset - if list.ItemsAvailable > len(list.Items) { - // TODO: support paging - log.Printf("Warning! %d containers are available but only received %d, paged requests are not yet supported, some containers may be ignored.", - list.ItemsAvailable, - len(list.Items)) + // This list variable must be a new one declared + // inside the loop: otherwise, items in the API + // response would get deep-merged into the items + // loaded in previous iterations. + var list arvados.ContainerList + + err := d.Arv.List("containers", params, &list) + if err != nil { + d.Logger.Warnf("error getting list of containers: %q", err) + return false + } + d.checkListForUpdates(list.Items, todo) + offset += len(list.Items) + if len(list.Items) == 0 || itemsAvailable <= offset { + return true + } } +} +func (d *Dispatcher) checkListForUpdates(containers []arvados.Container, todo map[string]*runTracker) { d.mtx.Lock() defer d.mtx.Unlock() - if d.running == nil { - d.running = make(map[string]*runTracker) + if d.trackers == nil { + d.trackers = make(map[string]*runTracker) } - for _, c := range list.Items { - tracker, running := d.running[c.UUID] + for _, c := range containers { + tracker, alreadyTracking := d.trackers[c.UUID] + delete(todo, c.UUID) + if c.LockedByUUID != "" && c.LockedByUUID != d.auth.UUID { - log.Printf("debug: ignoring %s locked by %s", c.UUID, c.LockedByUUID) - } else if running { + d.Logger.Debugf("ignoring %s locked by %s", c.UUID, c.LockedByUUID) + } else if alreadyTracking { switch c.State { case Queued: tracker.close() @@ -162,18 +251,18 @@ func (d *Dispatcher) checkForUpdates(filters [][]interface{}) { } err := d.lock(c.UUID) if err != nil { - log.Printf("debug: error locking container %s: %s", c.UUID, err) + d.Logger.Warnf("error locking container %s: %s", c.UUID, err) break } c.State = Locked - d.running[c.UUID] = d.start(c) + d.trackers[c.UUID] = d.start(c) case Locked, Running: if !d.throttle.Check(c.UUID) { break } - d.running[c.UUID] = d.start(c) + d.trackers[c.UUID] = d.start(c) case Cancelled, Complete: - tracker.close() + // no-op (we already stopped monitoring) } } } @@ -186,7 +275,7 @@ func (d *Dispatcher) UpdateState(uuid string, state arvados.ContainerState) erro "container": arvadosclient.Dict{"state": state}, }, nil) if err != nil { - log.Printf("Error updating container %s to state %q: %s", uuid, state, err) + d.Logger.Warnf("error updating container %s to state %q: %s", uuid, state, err) } return err } @@ -201,9 +290,47 @@ func (d *Dispatcher) Unlock(uuid string) error { return d.Arv.Call("POST", "containers", uuid, "unlock", nil, nil) } +// TrackContainer ensures a tracker is running for the given UUID, +// regardless of the current state of the container (except: if the +// container is locked by a different dispatcher, a tracker will not +// be started). If the container is not in Locked or Running state, +// the new tracker will close down immediately. +// +// This allows the dispatcher to put its own RunContainer func into a +// cleanup phase (for example, to kill local processes created by a +// prevous dispatch process that are still running even though the +// container state is final) without the risk of having multiple +// goroutines monitoring the same UUID. +func (d *Dispatcher) TrackContainer(uuid string) error { + var cntr arvados.Container + err := d.Arv.Call("GET", "containers", uuid, "", nil, &cntr) + if err != nil { + return err + } + if cntr.LockedByUUID != "" && cntr.LockedByUUID != d.auth.UUID { + return nil + } + + d.mtx.Lock() + defer d.mtx.Unlock() + if _, alreadyTracking := d.trackers[uuid]; alreadyTracking { + return nil + } + if d.trackers == nil { + d.trackers = make(map[string]*runTracker) + } + d.trackers[uuid] = d.start(cntr) + switch cntr.State { + case Queued, Cancelled, Complete: + d.trackers[uuid].close() + } + return nil +} + type runTracker struct { closing bool updates chan arvados.Container + logger Logger } func (tracker *runTracker) close() { @@ -219,7 +346,7 @@ func (tracker *runTracker) update(c arvados.Container) { } select { case <-tracker.updates: - log.Printf("debug: runner is handling updates slowly, discarded previous update for %s", c.UUID) + tracker.logger.Debugf("runner is handling updates slowly, discarded previous update for %s", c.UUID) default: } tracker.updates <- c