squall

In-memory stand-ins for cloud APIs that fail on cue, for testing code that must survive partial failure.

Code that creates cloud resources has to cope with calls that fail after they took effect: an insert that succeeded but whose response was lost, an operation that ended in an error, a retry that finds the resource already there. Those paths are where leaked resources and stuck records come from, and they are hard to reach against a real cloud. A squall server answers the same API calls the official client libraries make, keeps the resources in memory, and fails exactly the calls you choose, or a seeded random share of them, so a failure replays from its seed.

Packages

Package	What it does
fault	Decides which calls fail and how: rejected, applied with the response lost, applied with an error, an operation that ends in an error, or slow. Scripted rules and a seeded random rate.
aws/ec2	Amazon EC2: instances (run with client tokens, describe with filters and pages, start, stop, terminate, change type), images (describe, register from snapshots, deregister), instance types, EBS volumes (create, describe, attach, detach, delete), EBS snapshots (create, describe with filters and pages, delete), placement groups, tags, Elastic IP addresses (allocate, describe, release, associate with an interface or instance, disassociate) and network interfaces (create with client tokens, describe, delete).
aws/s3	Amazon S3: general purpose buckets (create, head, list with pages and filters, delete), their tags, versioning, policy and CORS rules, and objects (put, delete one or many, list versions and delete markers).
azure/arm	Azure Resource Manager: resource groups (get, create or update, list with tag filters, delete with everything in them that nothing outside uses), virtual machines (create or update, get, list, instance view, deallocate, start, delete), Uniform scale sets (create or update, update, get, list, delete, delete instances) and their VMs, network security groups, network interfaces and public IP addresses, the IP configurations of scale set VMs, private DNS zones, virtual network links and A record sets, and long-running operations to poll.
gcp/compute	Compute Engine: instances (bulkInsert, insert, get, list, aggregatedList, delete, stop, start, setLabels, setMachineType, addAccessConfig, attachDisk, detachDisk), disks (get, list, aggregatedList, insert, delete), snapshots (insert, get, list, delete), zones, regional addresses (get, list, insert, delete, setLabels), resource policies, instance templates, routes, and zone, region and global operations (get, wait).
gcp/storage	Cloud Storage: buckets (insert, get, list, patch labels, versioning and CORS, delete) and objects (multipart and media uploads, list with versions, delete by generation).

Coverage grows with use: each stand-in implements the methods someone needed, answers anything else with 404, and documents how faithful it is. Contributions of more methods and services are welcome.

Example

faults := fault.New().Add(fault.Rule{
	Method: "compute.addresses.insert",
	Nth:    1,
	Fault:  fault.LoseResponse(), // takes effect, but the caller never hears back
})
s := compute.New(compute.WithFaults(faults))
srv := httptest.NewServer(s)
defer srv.Close()

client, _ := gcompute.NewAddressesRESTClient(ctx,
	option.WithEndpoint(srv.URL), option.WithoutAuthentication())

// Run the code under test with client, then check what it left behind:
for _, a := range s.Addresses() {
	fmt.Println(a.Name, a.Address, a.Labels)
}

For a property test, inject random faults and check that every seed converges:

for seed := range uint64(100) {
	faults := fault.New(fault.Seed(seed)).Random(0.3,
		fault.Unavailable(), fault.LoseResponse(), fault.FailAfter(500))
	// ... run the code under test against compute.New(compute.WithFaults(faults))
	// and on failure report the seed and faults.Injected().
}

Faithfulness

Each stand-in follows the real API's requests, response shapes and error codes for the methods it covers, and is tested through the official Go client library. Where it simplifies, the package documentation says so. For Compute Engine:

• Changes to instances, disks and snapshots finish before the call that makes them returns, so the SDK's Operation.Wait returns at once. WithSettleAfter(n) keeps them in progress, with their operation RUNNING, for n more reads of the resource or its operation, or until the test calls Settle: a new instance is PROVISIONING and then STAGING, a stopping or deleted one STOPPING, a starting one STAGING, a new snapshot CREATING and a deleted one DELETING. An operations.wait finishes the operation it waits for. A change to an instance that has another in progress finishes that one first. Every other change finishes at once.
• A call that repeats an earlier call's requestId in the same project, for the same method, gets that call's operation back without taking effect again, as the real API does for retries. The server keeps the last 10,000 operations.
• Images are registered by the test with PutImage, and an image family names the newest image in it. Snapshots are taken of a READY disk with snapshots.insert, or registered with PutSnapshot; they record the disk's size, not its data, and a disk from one that is not READY fails with resourceNotReady. Machine and disk types are any well-formed name in the call's zone. Zones come from a fixed set per region, which WithZones changes.
• Networks, subnetworks, service accounts and accelerators are not checked. Fields the stand-in only records keep the JSON the caller sent, and fields it does not know are rejected with 400. Instance group managers are not supported: their aggregated list is always empty.
• A bulkInsert skips each instance whose name, or a disk it would create, already exists, and creates the rest when they number at least minCount; its operation succeeds, with createdVmCount and failedToCreateVmCount counting the two, and lists no warning for what it skipped. With fewer than minCount, its operation fails with VM_MIN_COUNT_NOT_REACHED and a RESOURCE_ALREADY_EXISTS error per conflict, and nothing is created. It takes no fixed IP addresses or disk names, and ignores locationPolicy.
• List filters support field = value and field != value comparisons joined by AND, OR and parentheses (OR binds tighter than AND), and the regular expressions field eq re and field ne re, which must match the whole field. Lists are ordered by name. Anything else, such as a wildcard or orderBy=creationTimestamp desc, is rejected with 400 rather than ignored. An aggregatedList lists only the scopes that have items.
• IP addresses come from 100.64.0.0/10 (external) and 10.0.0.0/8 (internal), never from routable ranges. An external address is unique across every project and region; an internal one is unique within its region, which stands in for its subnetwork. An instance's ephemeral external address is released when it stops; a reserved one it takes is IN_USE until it goes, and stays with it until then if the address is deleted.
• A fault rule's Resource for an insert is the resource being created, so a rule can target the insert of one name; for a bulkInsert, the first instance it names.

For Amazon EC2:

• Signatures are not checked. The region comes from the SigV4 credential scope, and each region keeps its own resources.
• Subnet and security group IDs and key pair names are not validated, and there are no VPCs, default or otherwise: RunInstances needs a subnet or a network interface. Images are registered by the test with PutImage, or with RegisterImage from snapshots the stand-in holds; instance types come from a small built-in set or PutInstanceType.
• Snapshots record the volume's size and encryption, not its data, and every one is the caller's own: DescribeSnapshots with an owner other than self lists none. DeleteSnapshot fails with InvalidSnapshot.InUse while a registered image maps the snapshot.
• RunInstances records metadata, private DNS name and CPU options, and the primary interface's type (interface or efa; with an existing interface, only that interface's type) and network card index (only 0, the same launch as leaving it out); DescribeInstances returns them, with metadata options pending until the instance first settles. EFA and AMD SEV-SNP are not checked against the instance type. CPU options take up to the type's default cores and threads per core, which DescribeInstanceTypes reports, and any core count from 1 up, where EC2 accepts only the counts it lists for each type.
• Instances, volumes and snapshots move through EC2's states when describes see them. By default the first DescribeInstances, DescribeVolumes or DescribeSnapshots after a change shows the state it settles in, so the SDK's waiters return at once; WithSettleAfter holds them in transition for more describes, or until the test calls Settle. Describes are consistent at once, not eventually.
• Terminated instances stay visible, as in EC2, until the test calls ForgetInstance. Termination deletes the interface and volumes marked to delete on termination and releases the rest.
• A CreateNetworkInterface, RunInstances or CreateVolume that repeats a ClientToken with the same parameters gets what the first call created, as the real API does for retries; with other parameters it fails with IdempotentParameterMismatch.
• Filters take EC2's * and ? wildcards. Filter names and parameters the stand-in does not support are rejected with 400 rather than ignored.
• Public IP addresses come from 100.64.0.0/10 and private ones from 10.0.0.0/8. There is no address quota unless the test sets one.
• EC2 has no long-running operations, so a failed operation fails the call before it takes effect. A fault's reason is the EC2 error code it answers with, and the generic ones become InternalError, Unavailable and RequestLimitExceeded. A fault rule's Resource is the region and ID, such as us-east-1/eipalloc-…, and for a create, the resource being created (the first instance, for RunInstances). A rejected call has no effect, so a rejected describe does not move an instance along.

For Amazon S3:

• Requests are path-style, which the SDK uses for an IP endpoint such as httptest's. Signatures are not checked; the region comes from the SigV4 credential scope. A call on a bucket in another region fails with 301 PermanentRedirect and x-amz-bucket-region, as S3 answers it.
• There is one account. PutForeignBucket holds a name for another one, so creating it fails with BucketAlreadyExists and calls on it with AccessDenied (403 for HeadBucket). Creating an owned bucket again succeeds in us-east-1 and fails with BucketAlreadyOwnedByYou elsewhere.
• ListBuckets with any parameter pages (10,000 buckets by default) and reports each bucket's region, as S3 does; with none it lists every bucket without regions. ListObjectVersions orders by key and then newest first, and its markers are positions, so the next page still follows one whose versions were just deleted. DeleteBucket fails with BucketNotEmpty while any version or delete marker remains.
• Objects keep their size and ETag, not their data. Checksums are not verified, and multipart uploads, copies, object tags and lifecycle are not supported. Policies and CORS rules are recorded, not enforced. S3 documents in-progress multipart uploads as blocking DeleteBucket only for directory buckets.
• A fault rule's Resource is the bucket name, or bucket/key for an object call. A lost HeadBucket breaks off inside its status line, since a sent header would read as an answer.

For Cloud Storage:

• Point the client at srv.URL + "/storage/v1/", or set STORAGE_EMULATOR_HOST. There is one caller, which reaches every project; PutForeignBucket holds a name it cannot reach, so inserting it fails with 409 conflict and calls on it with 403 forbidden. A repeated insert of the caller's own bucket fails with 409, as the client's retry of a create that took effect does.
• Bucket names and labels follow the API's rules; locations are upper-cased but not checked. Fields the server does not model, such as lifecycle rules or retention, are rejected with 400. A patch with a field set to null clears it, and a label set to null is deleted.
• Objects keep their size, MD5 and CRC32C, which an upload's checksums must match. With versioning on, overwrites and deletes keep the old generation as noncurrent, and a bucket with any generation left cannot be deleted. Page tokens are positions, so a listing survives deletes between pages. Resumable uploads, reads, IAM policies and HMAC keys are not supported.
• A fault rule's Resource is the path below /storage/v1/, such as b/my-bucket/o/my-object, and for an insert the bucket or object it makes.

For Azure Resource Manager:

• Long-running operations finish within the call that starts them, so the SDK's pollers are done at once. WithSettleAfter(n) answers with an Azure-AsyncOperation URL instead (and a Location URL for actions and deletes, the only one for a resource group deletion), keeping the resource Creating, Updating or Deleting for n more reads of it or its operation, or until the test calls Settle. Polls carry a retry-after-ms of 10. A fault in FailOperation mode ends the operation in its error without the call taking effect.
• PUT and DELETE are idempotent, as in Azure: a retried create answers 200 with the resource, and deleting what is not there answers 204. A change to a resource with another in progress is answered as its provider answers it: network resources fail it with 429 RetryableError and virtual network links with 409 Conflict, counting the call as a read of that change, and a VM or scale set change preempts the other, which ends Canceled with OperationPreempted, or fails with 409 OperationNotAllowed if that one is a deletion. A change in a resource group being deleted fails with ResourceGroupBeingDeleted.
• Deleting a resource group deletes what nothing outside it uses. When something outside still uses a resource in it, such as a VM in another group attached to its network interface, the rest is deleted and the operation fails with 409 ResourceGroupDeletionBlocked, leaving the group with the resources in use. Each resource left is a detail whose message is the error deleting it alone answers, such as NicInUse, as Azure reports it; the status, and that the rest is deleted first, are inferred.
• Resource names, subscription IDs, locations and tags are not validated, and names match without regard to case. Fields it does not know are rejected with InvalidRequestContent; values it does not model, such as a Flexible scale set or a VM with networkInterfaceConfigurations, with BadRequest.
• Virtual networks and subnets are not modeled; their IDs are recorded. Private addresses come from 10.0.0.0/8, unique within a subnet ID, and public ones from 100.64.0.0/10, assigned when created; WithPrivateRange and WithPublicRange change the ranges. Once a range is used up, a change that needs an address fails with SubnetIsFull or PublicIPCountLimitReached. OS disks get names and IDs but are not served.
• Scale set VMs are numbered from 0 without reuse; raising sku.capacity creates them with their interfaces and addresses, and lowering it deletes the newest. A PUT to an existing scale set that leaves out sku keeps the capacity. One that gives a sku without a capacity fails, as Azure answers one that changes the model, unless it repeats the create of a scale set made without a capacity with nothing changed since: then it succeeds and changes nothing, so a retry after a lost response converges, which is inferred from ARM PUTs being idempotent. A new scale set needs a sku with a name. One created without a capacity starts with no VMs, as a Flexible one does in Azure; for Uniform, both are inferred.
• Resource SKUs list none, so availability checks see no restrictions.
• Fault method names are the Azure RBAC action, such as Microsoft.Compute/virtualMachines/write, and a rule's Resource is the resource ID in the call's path.

Requirements

Go 1.27 or later. The module requires nothing beyond the standard library; the tests that use the official client libraries live in the separate sdktest module.

Contributing

See CONTRIBUTING.md. Security issues: SECURITY.md.

License

Apache 2.0.