1 // Copyright (C) The Arvados Authors. All rights reserved.
3 // SPDX-License-Identifier: AGPL-3.0
22 "git.arvados.org/arvados.git/lib/cloud"
23 "git.arvados.org/arvados.git/sdk/go/arvados"
24 "github.com/aws/aws-sdk-go/aws"
25 "github.com/aws/aws-sdk-go/aws/awserr"
26 "github.com/aws/aws-sdk-go/aws/credentials"
27 "github.com/aws/aws-sdk-go/aws/credentials/ec2rolecreds"
28 "github.com/aws/aws-sdk-go/aws/ec2metadata"
29 "github.com/aws/aws-sdk-go/aws/request"
30 "github.com/aws/aws-sdk-go/aws/session"
31 "github.com/aws/aws-sdk-go/service/ec2"
32 "github.com/prometheus/client_golang/prometheus"
33 "github.com/sirupsen/logrus"
34 "golang.org/x/crypto/ssh"
37 // Driver is the ec2 implementation of the cloud.Driver interface.
38 var Driver = cloud.DriverFunc(newEC2InstanceSet)
41 throttleDelayMin = time.Second
42 throttleDelayMax = time.Minute
45 type ec2InstanceSetConfig struct {
47 SecretAccessKey string
49 SecurityGroupIDs arvados.StringSet
50 SubnetID sliceOrSingleString
54 IAMInstanceProfile string
55 SpotPriceUpdateInterval arvados.Duration
58 type sliceOrSingleString []string
60 // UnmarshalJSON unmarshals an array of strings, and also accepts ""
61 // as [], and "foo" as ["foo"].
62 func (ss *sliceOrSingleString) UnmarshalJSON(data []byte) error {
65 } else if data[0] == '[' {
67 err := json.Unmarshal(data, &slice)
78 err := json.Unmarshal(data, &str)
91 type ec2Interface interface {
92 DescribeKeyPairs(input *ec2.DescribeKeyPairsInput) (*ec2.DescribeKeyPairsOutput, error)
93 ImportKeyPair(input *ec2.ImportKeyPairInput) (*ec2.ImportKeyPairOutput, error)
94 RunInstances(input *ec2.RunInstancesInput) (*ec2.Reservation, error)
95 DescribeInstances(input *ec2.DescribeInstancesInput) (*ec2.DescribeInstancesOutput, error)
96 DescribeInstanceStatusPages(input *ec2.DescribeInstanceStatusInput, fn func(*ec2.DescribeInstanceStatusOutput, bool) bool) error
97 DescribeSpotPriceHistoryPages(input *ec2.DescribeSpotPriceHistoryInput, fn func(*ec2.DescribeSpotPriceHistoryOutput, bool) bool) error
98 CreateTags(input *ec2.CreateTagsInput) (*ec2.CreateTagsOutput, error)
99 TerminateInstances(input *ec2.TerminateInstancesInput) (*ec2.TerminateInstancesOutput, error)
102 type ec2InstanceSet struct {
103 ec2config ec2InstanceSetConfig
104 currentSubnetIDIndex int32
105 instanceSetID cloud.InstanceSetID
106 logger logrus.FieldLogger
109 keys map[string]string
110 throttleDelayCreate atomic.Value
111 throttleDelayInstances atomic.Value
113 prices map[priceKey][]cloud.InstancePrice
114 pricesLock sync.Mutex
115 pricesUpdated map[priceKey]time.Time
118 func newEC2InstanceSet(config json.RawMessage, instanceSetID cloud.InstanceSetID, _ cloud.SharedResourceTags, logger logrus.FieldLogger, reg *prometheus.Registry) (prv cloud.InstanceSet, err error) {
119 instanceSet := &ec2InstanceSet{
120 instanceSetID: instanceSetID,
123 err = json.Unmarshal(config, &instanceSet.ec2config)
128 sess, err := session.NewSession()
132 // First try any static credentials, fall back to an IAM instance profile/role
133 creds := credentials.NewChainCredentials(
134 []credentials.Provider{
135 &credentials.StaticProvider{Value: credentials.Value{AccessKeyID: instanceSet.ec2config.AccessKeyID, SecretAccessKey: instanceSet.ec2config.SecretAccessKey}},
136 &ec2rolecreds.EC2RoleProvider{Client: ec2metadata.New(sess)},
139 awsConfig := aws.NewConfig().WithCredentials(creds).WithRegion(instanceSet.ec2config.Region)
140 instanceSet.client = ec2.New(session.Must(session.NewSession(awsConfig)))
141 instanceSet.keys = make(map[string]string)
142 if instanceSet.ec2config.EBSVolumeType == "" {
143 instanceSet.ec2config.EBSVolumeType = "gp2"
145 return instanceSet, nil
148 func awsKeyFingerprint(pk ssh.PublicKey) (md5fp string, sha1fp string, err error) {
149 // AWS key fingerprints don't use the usual key fingerprint
150 // you get from ssh-keygen or ssh.FingerprintLegacyMD5()
151 // (you can get that from md5.Sum(pk.Marshal())
153 // AWS uses the md5 or sha1 of the PKIX DER encoding of the
154 // public key, so calculate those fingerprints here.
160 if err := ssh.Unmarshal(pk.Marshal(), &rsaPub); err != nil {
161 return "", "", fmt.Errorf("agent: Unmarshal failed to parse public key: %v", err)
163 rsaPk := rsa.PublicKey{
164 E: int(rsaPub.E.Int64()),
167 pkix, _ := x509.MarshalPKIXPublicKey(&rsaPk)
168 md5pkix := md5.Sum([]byte(pkix))
169 sha1pkix := sha1.Sum([]byte(pkix))
172 for i := 0; i < len(md5pkix); i++ {
173 md5fp += fmt.Sprintf(":%02x", md5pkix[i])
175 for i := 0; i < len(sha1pkix); i++ {
176 sha1fp += fmt.Sprintf(":%02x", sha1pkix[i])
178 return md5fp[1:], sha1fp[1:], nil
181 func (instanceSet *ec2InstanceSet) Create(
182 instanceType arvados.InstanceType,
183 imageID cloud.ImageID,
184 newTags cloud.InstanceTags,
185 initCommand cloud.InitCommand,
186 publicKey ssh.PublicKey) (cloud.Instance, error) {
188 ec2tags := []*ec2.Tag{}
189 for k, v := range newTags {
190 ec2tags = append(ec2tags, &ec2.Tag{
192 Value: aws.String(v),
197 for sg := range instanceSet.ec2config.SecurityGroupIDs {
198 groups = append(groups, sg)
201 rii := ec2.RunInstancesInput{
202 ImageId: aws.String(string(imageID)),
203 InstanceType: &instanceType.ProviderType,
204 MaxCount: aws.Int64(1),
205 MinCount: aws.Int64(1),
207 NetworkInterfaces: []*ec2.InstanceNetworkInterfaceSpecification{
209 AssociatePublicIpAddress: aws.Bool(false),
210 DeleteOnTermination: aws.Bool(true),
211 DeviceIndex: aws.Int64(0),
212 Groups: aws.StringSlice(groups),
214 DisableApiTermination: aws.Bool(false),
215 InstanceInitiatedShutdownBehavior: aws.String("terminate"),
216 TagSpecifications: []*ec2.TagSpecification{
218 ResourceType: aws.String("instance"),
221 UserData: aws.String(base64.StdEncoding.EncodeToString([]byte("#!/bin/sh\n" + initCommand + "\n"))),
224 if publicKey != nil {
225 keyname, err := instanceSet.getKeyName(publicKey)
229 rii.KeyName = &keyname
232 if instanceType.AddedScratch > 0 {
233 rii.BlockDeviceMappings = []*ec2.BlockDeviceMapping{{
234 DeviceName: aws.String("/dev/xvdt"),
235 Ebs: &ec2.EbsBlockDevice{
236 DeleteOnTermination: aws.Bool(true),
237 VolumeSize: aws.Int64((int64(instanceType.AddedScratch) + (1<<30 - 1)) >> 30),
238 VolumeType: &instanceSet.ec2config.EBSVolumeType,
242 if instanceType.Preemptible {
243 rii.InstanceMarketOptions = &ec2.InstanceMarketOptionsRequest{
244 MarketType: aws.String("spot"),
245 SpotOptions: &ec2.SpotMarketOptions{
246 InstanceInterruptionBehavior: aws.String("terminate"),
247 MaxPrice: aws.String(fmt.Sprintf("%v", instanceType.Price)),
251 if instanceSet.ec2config.IAMInstanceProfile != "" {
252 rii.IamInstanceProfile = &ec2.IamInstanceProfileSpecification{
253 Name: aws.String(instanceSet.ec2config.IAMInstanceProfile),
257 var rsv *ec2.Reservation
259 subnets := instanceSet.ec2config.SubnetID
260 currentSubnetIDIndex := int(atomic.LoadInt32(&instanceSet.currentSubnetIDIndex))
261 for tryOffset := 0; ; tryOffset++ {
263 if len(subnets) > 0 {
264 tryIndex = (currentSubnetIDIndex + tryOffset) % len(subnets)
265 rii.NetworkInterfaces[0].SubnetId = aws.String(subnets[tryIndex])
267 rsv, err = instanceSet.client.RunInstances(&rii)
268 if isErrorSubnetSpecific(err) &&
269 tryOffset < len(subnets)-1 {
270 instanceSet.logger.WithError(err).WithField("SubnetID", subnets[tryIndex]).
271 Warn("RunInstances failed, trying next subnet")
274 // Succeeded, or exhausted all subnets, or got a
275 // non-subnet-related error.
277 // We intentionally update currentSubnetIDIndex even
278 // in the non-retryable-failure case here to avoid a
279 // situation where successive calls to Create() keep
280 // returning errors for the same subnet (perhaps
281 // "subnet full") and never reveal the errors for the
282 // other configured subnets (perhaps "subnet ID
284 atomic.StoreInt32(&instanceSet.currentSubnetIDIndex, int32(tryIndex))
287 err = wrapError(err, &instanceSet.throttleDelayCreate)
292 provider: instanceSet,
293 instance: rsv.Instances[0],
297 func (instanceSet *ec2InstanceSet) getKeyName(publicKey ssh.PublicKey) (string, error) {
298 instanceSet.keysMtx.Lock()
299 defer instanceSet.keysMtx.Unlock()
300 md5keyFingerprint, sha1keyFingerprint, err := awsKeyFingerprint(publicKey)
302 return "", fmt.Errorf("Could not make key fingerprint: %v", err)
304 if keyname, ok := instanceSet.keys[md5keyFingerprint]; ok {
307 keyout, err := instanceSet.client.DescribeKeyPairs(&ec2.DescribeKeyPairsInput{
308 Filters: []*ec2.Filter{{
309 Name: aws.String("fingerprint"),
310 Values: []*string{&md5keyFingerprint, &sha1keyFingerprint},
314 return "", fmt.Errorf("Could not search for keypair: %v", err)
316 if len(keyout.KeyPairs) > 0 {
317 return *(keyout.KeyPairs[0].KeyName), nil
319 keyname := "arvados-dispatch-keypair-" + md5keyFingerprint
320 _, err = instanceSet.client.ImportKeyPair(&ec2.ImportKeyPairInput{
322 PublicKeyMaterial: ssh.MarshalAuthorizedKey(publicKey),
325 return "", fmt.Errorf("Could not import keypair: %v", err)
327 instanceSet.keys[md5keyFingerprint] = keyname
331 func (instanceSet *ec2InstanceSet) Instances(tags cloud.InstanceTags) (instances []cloud.Instance, err error) {
332 var filters []*ec2.Filter
333 for k, v := range tags {
334 filters = append(filters, &ec2.Filter{
335 Name: aws.String("tag:" + k),
336 Values: []*string{aws.String(v)},
340 dii := &ec2.DescribeInstancesInput{Filters: filters}
342 dio, err := instanceSet.client.DescribeInstances(dii)
343 err = wrapError(err, &instanceSet.throttleDelayInstances)
348 for _, rsv := range dio.Reservations {
349 for _, inst := range rsv.Instances {
350 if *inst.State.Name != "shutting-down" && *inst.State.Name != "terminated" {
351 instances = append(instances, &ec2Instance{
352 provider: instanceSet,
355 if aws.StringValue(inst.InstanceLifecycle) == "spot" {
361 if dio.NextToken == nil {
364 dii.NextToken = dio.NextToken
366 if needAZs && instanceSet.ec2config.SpotPriceUpdateInterval > 0 {
367 az := map[string]string{}
368 err := instanceSet.client.DescribeInstanceStatusPages(&ec2.DescribeInstanceStatusInput{
369 IncludeAllInstances: aws.Bool(true),
370 }, func(page *ec2.DescribeInstanceStatusOutput, lastPage bool) bool {
371 for _, ent := range page.InstanceStatuses {
372 az[*ent.InstanceId] = *ent.AvailabilityZone
377 instanceSet.logger.Warnf("error getting instance statuses: %s", err)
379 for _, inst := range instances {
380 inst := inst.(*ec2Instance)
381 inst.availabilityZone = az[*inst.instance.InstanceId]
383 instanceSet.updateSpotPrices(instances)
385 return instances, err
388 type priceKey struct {
391 availabilityZone string
394 // Refresh recent spot instance pricing data for the given instances,
395 // unless we already have recent pricing data for all relevant types.
396 func (instanceSet *ec2InstanceSet) updateSpotPrices(instances []cloud.Instance) {
397 if len(instances) == 0 {
401 instanceSet.pricesLock.Lock()
402 defer instanceSet.pricesLock.Unlock()
403 if instanceSet.prices == nil {
404 instanceSet.prices = map[priceKey][]cloud.InstancePrice{}
405 instanceSet.pricesUpdated = map[priceKey]time.Time{}
408 updateTime := time.Now()
409 staleTime := updateTime.Add(-instanceSet.ec2config.SpotPriceUpdateInterval.Duration())
411 allTypes := map[string]bool{}
413 for _, inst := range instances {
414 ec2inst := inst.(*ec2Instance).instance
415 if aws.StringValue(ec2inst.InstanceLifecycle) == "spot" {
417 instanceType: *ec2inst.InstanceType,
419 availabilityZone: inst.(*ec2Instance).availabilityZone,
421 if instanceSet.pricesUpdated[pk].Before(staleTime) {
424 allTypes[*ec2inst.InstanceType] = true
430 var typeFilterValues []*string
431 for instanceType := range allTypes {
432 typeFilterValues = append(typeFilterValues, aws.String(instanceType))
434 // Get 3x update interval worth of pricing data. (Ideally the
435 // AWS API would tell us "we have shown you all of the price
436 // changes up to time T", but it doesn't, so we'll just ask
437 // for 3 intervals worth of data on each update, de-duplicate
438 // the data points, and not worry too much about occasionally
439 // missing some data points when our lookups fail twice in a
441 dsphi := &ec2.DescribeSpotPriceHistoryInput{
442 StartTime: aws.Time(updateTime.Add(-3 * instanceSet.ec2config.SpotPriceUpdateInterval.Duration())),
443 Filters: []*ec2.Filter{
444 &ec2.Filter{Name: aws.String("instance-type"), Values: typeFilterValues},
445 &ec2.Filter{Name: aws.String("product-description"), Values: []*string{aws.String("Linux/UNIX")}},
448 err := instanceSet.client.DescribeSpotPriceHistoryPages(dsphi, func(page *ec2.DescribeSpotPriceHistoryOutput, lastPage bool) bool {
449 for _, ent := range page.SpotPriceHistory {
450 if ent.InstanceType == nil || ent.SpotPrice == nil || ent.Timestamp == nil {
454 price, err := strconv.ParseFloat(*ent.SpotPrice, 64)
460 instanceType: *ent.InstanceType,
462 availabilityZone: *ent.AvailabilityZone,
464 instanceSet.prices[pk] = append(instanceSet.prices[pk], cloud.InstancePrice{
465 StartTime: *ent.Timestamp,
468 instanceSet.pricesUpdated[pk] = updateTime
473 instanceSet.logger.Warnf("error retrieving spot instance prices: %s", err)
476 expiredTime := updateTime.Add(-64 * instanceSet.ec2config.SpotPriceUpdateInterval.Duration())
477 for pk, last := range instanceSet.pricesUpdated {
478 if last.Before(expiredTime) {
479 delete(instanceSet.pricesUpdated, pk)
480 delete(instanceSet.prices, pk)
483 for pk, prices := range instanceSet.prices {
484 instanceSet.prices[pk] = cloud.NormalizePriceHistory(prices)
488 func (instanceSet *ec2InstanceSet) Stop() {
491 type ec2Instance struct {
492 provider *ec2InstanceSet
493 instance *ec2.Instance
494 availabilityZone string // sometimes available for spot instances
497 func (inst *ec2Instance) ID() cloud.InstanceID {
498 return cloud.InstanceID(*inst.instance.InstanceId)
501 func (inst *ec2Instance) String() string {
502 return *inst.instance.InstanceId
505 func (inst *ec2Instance) ProviderType() string {
506 return *inst.instance.InstanceType
509 func (inst *ec2Instance) SetTags(newTags cloud.InstanceTags) error {
510 var ec2tags []*ec2.Tag
511 for k, v := range newTags {
512 ec2tags = append(ec2tags, &ec2.Tag{
514 Value: aws.String(v),
518 _, err := inst.provider.client.CreateTags(&ec2.CreateTagsInput{
519 Resources: []*string{inst.instance.InstanceId},
526 func (inst *ec2Instance) Tags() cloud.InstanceTags {
527 tags := make(map[string]string)
529 for _, t := range inst.instance.Tags {
530 tags[*t.Key] = *t.Value
536 func (inst *ec2Instance) Destroy() error {
537 _, err := inst.provider.client.TerminateInstances(&ec2.TerminateInstancesInput{
538 InstanceIds: []*string{inst.instance.InstanceId},
543 func (inst *ec2Instance) Address() string {
544 if inst.instance.PrivateIpAddress != nil {
545 return *inst.instance.PrivateIpAddress
550 func (inst *ec2Instance) RemoteUser() string {
551 return inst.provider.ec2config.AdminUsername
554 func (inst *ec2Instance) VerifyHostKey(ssh.PublicKey, *ssh.Client) error {
555 return cloud.ErrNotImplemented
558 // PriceHistory returns the price history for this specific instance.
560 // AWS documentation is elusive about whether the hourly cost of a
561 // given spot instance changes as the current spot price changes for
562 // the corresponding instance type and availability zone. Our
563 // implementation assumes the answer is yes, based on the following
566 // https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/spot-requests.html
567 // says: "After your Spot Instance is running, if the Spot price rises
568 // above your maximum price, Amazon EC2 interrupts your Spot
569 // Instance." (This doesn't address what happens when the spot price
570 // rises *without* exceeding your maximum price.)
572 // https://docs.aws.amazon.com/whitepapers/latest/cost-optimization-leveraging-ec2-spot-instances/how-spot-instances-work.html
573 // says: "You pay the Spot price that's in effect, billed to the
574 // nearest second." (But it's not explicitly stated whether "the price
575 // in effect" changes over time for a given instance.)
577 // The same page also says, in a discussion about the effect of
578 // specifying a maximum price: "Note that you never pay more than the
579 // Spot price that is in effect when your Spot Instance is running."
580 // (The use of the phrase "is running", as opposed to "was launched",
581 // hints that pricing is dynamic.)
582 func (inst *ec2Instance) PriceHistory(instType arvados.InstanceType) []cloud.InstancePrice {
583 inst.provider.pricesLock.Lock()
584 defer inst.provider.pricesLock.Unlock()
585 // Note updateSpotPrices currently populates
586 // inst.provider.prices only for spot instances, so if
587 // spot==false here, we will return no data.
589 instanceType: *inst.instance.InstanceType,
590 spot: aws.StringValue(inst.instance.InstanceLifecycle) == "spot",
591 availabilityZone: inst.availabilityZone,
593 var prices []cloud.InstancePrice
594 for _, price := range inst.provider.prices[pk] {
595 // ceil(added scratch space in GiB)
596 gib := (instType.AddedScratch + 1<<30 - 1) >> 30
597 monthly := inst.provider.ec2config.EBSPrice * float64(gib)
598 hourly := monthly / 30 / 24
599 price.Price += hourly
600 prices = append(prices, price)
605 type rateLimitError struct {
607 earliestRetry time.Time
610 func (err rateLimitError) EarliestRetry() time.Time {
611 return err.earliestRetry
614 var isCodeCapacity = map[string]bool{
615 "InstanceLimitExceeded": true,
616 "InsufficientAddressCapacity": true,
617 "InsufficientFreeAddressesInSubnet": true,
618 "InsufficientInstanceCapacity": true,
619 "InsufficientVolumeCapacity": true,
620 "MaxSpotInstanceCountExceeded": true,
621 "VcpuLimitExceeded": true,
624 // isErrorCapacity returns whether the error is to be throttled based on its code.
625 // Returns false if error is nil.
626 func isErrorCapacity(err error) bool {
627 if aerr, ok := err.(awserr.Error); ok && aerr != nil {
628 if _, ok := isCodeCapacity[aerr.Code()]; ok {
635 // isErrorSubnetSpecific returns true if the problem encountered by
636 // RunInstances might be avoided by trying a different subnet.
637 func isErrorSubnetSpecific(err error) bool {
638 aerr, ok := err.(awserr.Error)
643 return strings.Contains(code, "Subnet") ||
644 code == "InsufficientInstanceCapacity" ||
645 code == "InsufficientVolumeCapacity"
648 type ec2QuotaError struct {
652 func (er *ec2QuotaError) IsQuotaError() bool {
656 func wrapError(err error, throttleValue *atomic.Value) error {
657 if request.IsErrorThrottle(err) {
658 // Back off exponentially until an upstream call
659 // either succeeds or returns a non-throttle error.
660 d, _ := throttleValue.Load().(time.Duration)
661 d = d*3/2 + time.Second
662 if d < throttleDelayMin {
664 } else if d > throttleDelayMax {
667 throttleValue.Store(d)
668 return rateLimitError{error: err, earliestRetry: time.Now().Add(d)}
669 } else if isErrorCapacity(err) {
670 return &ec2QuotaError{err}
671 } else if err != nil {
672 throttleValue.Store(time.Duration(0))
675 throttleValue.Store(time.Duration(0))