fleet/docs/Contributing/reference/patterns-backend.md

Backend patterns

The backend software patterns that we follow in Fleet.

NOTE: There are always exceptions to the rules, but we try to follow these patterns as much as possible unless a specific use case calls for something else. These should be discussed within the team and documented before merging.

Table of Contents

• API Inputs
• Go
• MySQL
  • Timestamps
  • UUIDs
  • Say no to goqu
  • Data retention
  • Re-usable transactionable functions
• Specific features
  • GitOps

API Inputs

Input preprocessing and validation

Validate API inputs and return a 4XX status code if invalid. If you did not do authorization checking before failing validation, skip the authorization check with svc.authz.SkipAuthorization(ctx).

Inputs corresponding to sortable or indexed DB fields should be preprocessed (trim spaces, normalize Unicode, etc.). Use utility method fleet.Preprocess(input string) string. Backend sync where discussed.

Invalid inputs should NOT log a server error. Server errors should be reserved for unexpected/serious issues. InvalidArgumentError implements IsClientError method to indicate that it is a client error. Backend sync where discussed.

JSON unmarshaling

PATCH API calls often need to distinguish between a field being set to null and a field not being present in the JSON. Use the structs from optjson package to handle this. Backend sync where discussed. JSON unmarshaling article and example.

Go

Integer number types

Use int number type for general integer numbers. See Why does len() returned a signed value? for some context.

Exceptions:

• Database IDs
• Extra range of unsigned needed for a specific use case
• Specific performance/memory requirements

Unit testing

Use multiple hosts in unit tests and manual QA. For example, use a Windows VM and a Windows bare metal host when testing Windows profiles. Since our customers run Fleet on many hosts, we must be vigilant regarding multi-host use cases. Backend sync where discussed.

MySQL

Timestamps

Use high precision for all time fields. Precise timestamps make sure that we can accurately track when records were created and updated, keep records in order with a reliable sort, and speed up testing by not having to wait for the time to update. MySQL reference. Backend sync where discussed. Example:

CREATE TABLE `sample` (
  `id` INT UNSIGNED NOT NULL AUTO_INCREMENT,
  `created_at` TIMESTAMP(6) NOT NULL DEFAULT CURRENT_TIMESTAMP(6),
  `updated_at` TIMESTAMP(6) NOT NULL DEFAULT CURRENT_TIMESTAMP(6) ON UPDATE CURRENT_TIMESTAMP(6),
  PRIMARY KEY (`id`)
);

UUIDs

Use binary (or varbinary) data type for UUIDs. MySQL 8 has good support for UUIDs with UUID_TO_BIN and BIN_TO_UUID functions. If needed, add a virtual table to display UUID as string. Backend sync where discussed.

Benefits of binary UUIDs include:

• Smaller storage size
• Faster indexing/lookup

Say no to goqu

Do not use goqu; use MySQL queries directly. Searching for, understanding, and debugging direct MySQL queries is easier. If needing to modify an existing goqu query, try to rewrite it in MySQL. Backend sync where discussed.

Data retention

Sometimes we need data from rows that have been deleted from DB. For example, the activity feed may be retained forever, and it needs user info (or host info) that may not exist anymore. We need to keep this data in dedicated table(s). When a user row is deleted, a new entry with the same user.id is added to users_deleted table. The user info can be retrieved using ds.UserOrDeletedUserByID method.

Re-usable transactionable functions

Sometimes we want to encapsulate a piece of functionality in such a way that it can be use both independently and as part of a transaction. To do so, create a private function in the following way:

func myTransactionableFunction(ctx context.Context, tx sqlx.ExtContext, yourArgsHere any) error {
  // some setup, statements, etc...

  _, err := tx.ExecContext(ctx, stmt, args)
	if err != nil {
		return ctxerr.Wrap(ctx, err, "doing some stuff in a transaction")
	}
}

You can then use the function as a standalone call, like so

// *sqlx.DB implements the sqlx.ExtContext interface
err := myTransactionableFunction(ctx, ds.writer(ctx), myArgs)

or as part of a transaction, like so

func (ds *Datastore) MyDSMethodWithTransaction(ctx context.Context, yourArgsHere any) error {
	return ds.withRetryTxx(ctx, func(tx sqlx.ExtContext) error {
		return myTransactionableFunction(ctx, tx, yourArgsHere)
	})
}

See this commit for an example of this pattern.

Specific features

GitOps

GitOps documentation

fleetctl gitops was implemented on top of the existing fleetctl apply command. Now that apply no longer supports the newest features, we need to separate the code for the two commands.

Common issues and gotchas:

• Removing a setting. When a setting is removed from the YAML config file, the GitOps run should remove it from the server. Sometimes, the removal doesn't happen since apply did not work like this. Also, developers/QA may forget to test this case explicitly.
• Few integration tests. GitOps is a complex feature with an extensive state space because many settings interact. At the same time, setting up a test environment for GitOps is difficult. As we work on GitOps, we need to add more integration tests and develop testing utilities to make adding future integration tests easier.
• GitOps admin can define settings in default.yml, teams/team-name.yml, or teams/no-team.yml. Create unit tests for all these cases for features that support them.