errors.go

// Package terrors implements an error wrapping library.
//
// Terrors are used to provide context to an error, offering a stack trace and
// user defined error parameters.
//
// Terrors can be used to wrap any object that satisfies the error interface:
//
//	terr := terrors.Wrap(err, map[string]string{"context": "my_context"})
//
// Terrors can be instantiated directly:
//
//	err := terrors.New("not_found", "object not found", map[string]string{
//		"context": "my_context"
//	})
//
// Terrors offers built-in functions for instantiating Errors with common codes:
//
//	err := terrors.NotFound("config_file", "config file not found", map[string]string{
//		"context": my_context
//	})
package terrors

import (
	"fmt"
	"strings"

	"github.com/monzo/terrors/stack"
)

// Generic error codes. Each of these has their own constructor for convenience.
// You can use any string as a code, just use the `New` method.
// Warning: any new generic error code must be added to GenericErrorCodes.
const (
	ErrBadRequest         = "bad_request"
	ErrBadResponse        = "bad_response"
	ErrForbidden          = "forbidden"
	ErrInternalService    = "internal_service"
	ErrNotFound           = "not_found"
	ErrPreconditionFailed = "precondition_failed"
	ErrTimeout            = "timeout"
	ErrUnauthorized       = "unauthorized"
	ErrUnknown            = "unknown"
	ErrRateLimited        = "rate_limited"
)

// GenericErrorCodes is a list of all well known generic error codes.
var GenericErrorCodes = []string{
	ErrBadRequest,
	ErrBadResponse,
	ErrForbidden,
	ErrInternalService,
	ErrNotFound,
	ErrPreconditionFailed,
	ErrTimeout,
	ErrUnauthorized,
	ErrUnknown,
	ErrRateLimited,
}

var retryableCodes = []string{
	ErrInternalService,
	ErrTimeout,
	ErrUnknown,
	ErrRateLimited,
}

// Error is terror's error. It implements Go's error interface.
type Error struct {
	Code        string            `json:"code"`
	Message     string            `json:"message"`
	Params      map[string]string `json:"params"`
	StackFrames stack.Stack       `json:"stack"`

	// Exported for serialization, but you should use Retryable to read the value.
	IsRetryable *bool `json:"is_retryable"`

	// Exported for serialization, but you should use Unexpected to read the value.
	IsUnexpected *bool `json:"is_unexpected"`

	// Incremented each time the error is marshalled so that we can tell (approximately) how many services the error
	// has propagated through.  Higher level code can use this to influence decisions, for example it may only be
	// desirable to retry on an error that's only been marshalled once to avoid retries on top of retries... ad nauseam
	MarshalCount int `json:"marshal_count"`

	// When errors are marshalled certain information is lost (e.g. the 'cause').  This means if an error travels through
	// a number of services (and it's potentially augmented at each hop) that the core error message may be lost. The
	// history of an error is often a helpful debugging aid, so MessageChain is used to track this.
	MessageChain []string `json:"message_chain"`

	// Cause is the initial cause of this error, and will be populated
	// when using the Propagate function. This is intentionally not exported
	// so that we don't serialize causes and send them across process boundaries.
	// The cause refers to the cause of the error within a given process, and you
	// should not expect it to contain information about terrors from other downstream
	// processes.
	cause error
}

// Error returns a string message of the error.
// It will contain the code and error message. If there is a causal chain, the
// message from each error in the chain will be added to the output.
func (p *Error) Error() string {
	if p.cause == nil {
		// Not sure if the empty code/message cases actually happen, but to be safe, defer to
		// the 'old' error message if there is no cause present (i.e. we're not using
		// new wrapping functionality)
		return p.legacyErrString()
	}
	return fmt.Sprintf("%s: %s", p.Code, p.ErrorMessage())
}

// ErrorMessage returns a string message of the error.
// It will contain the error message, but not the code. If there is a causal
// chain, the message from each error in the chain will be added to the output.
func (p *Error) ErrorMessage() string {
	output := strings.Builder{}
	output.WriteString(p.Message)
	var next error = p.cause
	for next != nil {
		output.WriteString(": ")
		switch typed := next.(type) {
		case *Error:
			output.WriteString(typed.Message)
			next = typed.cause
		case error:
			output.WriteString(typed.Error())
			next = nil
		}
	}
	return output.String()
}

func (p *Error) legacyErrString() string {
	if p == nil {
		return ""
	}
	if p.Message == "" {
		return p.Code
	}
	if p.Code == "" {
		return p.Message
	}
	return fmt.Sprintf("%s: %s", p.Code, p.Message)
}

// Unwrap retruns the cause of the error. It may be nil.
func (p *Error) Unwrap() error {
	return p.cause
}

// StackTrace returns a slice of program counters taken from the stack frames.
// This adapts the terrors package to allow stacks to be reported to Sentry correctly.
func (p *Error) StackTrace() []uintptr {
	out := make([]uintptr, len(p.StackFrames))
	for i := 0; i < len(p.StackFrames); i++ {
		out[i] = p.StackFrames[i].PC
	}
	return out
}

// StackString formats the stacks from the terror chain as a string. If we
// encounter more than one terror in the chain with a stack frame, we'll print
// each one, separated by three hyphens on their own line.
func (p *Error) StackString() string {
	// 32,000 seems like a reasonable limit for a stack trace. Otherwise, we risk
	// overwhelming downstream systems.
	return StackStringWithMaxSize(p, 32000)
}

func StackStringWithMaxSize(p *Error, sizeLimit int) string {
	// if we run into this many causes, we've likely run into something absurd. Like
	// a self causing error.
	const maxCausalDepth = 1024
	var buffer strings.Builder
	terr := p
	var causalDepth int
outer:
	for terr != nil {
		if buffer.Len() != 0 && len(terr.StackFrames) > 0 {
			fmt.Fprintf(&buffer, "\n---")
		}
		for _, frame := range terr.StackFrames {
			// 10 seems like a reasonable estimate of how large the rest of the line would be.
			estimatedLineLen := len(frame.Filename) + len(frame.Method) + 16
			if estimatedLineLen+buffer.Len() > sizeLimit {
				break outer
			}
			fmt.Fprintf(&buffer, "\n  %s:%d in %s", frame.Filename, frame.Line, frame.Method)
		}

		if tcause, ok := terr.cause.(*Error); ok && causalDepth < maxCausalDepth {
			terr = tcause
			causalDepth += 1
		} else {
			break outer
		}
	}

	return buffer.String()
}

// VerboseString returns the error message, stack trace and params
func (p *Error) VerboseString() string {
	return fmt.Sprintf("%s\nParams: %+v\n%s", p.Error(), p.Params, p.StackString())
}

// Retryable determines whether the error was caused by an action which can be retried.
func (p *Error) Retryable() bool {
	if p.IsRetryable != nil {
		return *p.IsRetryable
	}
	for _, c := range retryableCodes {
		if PrefixMatches(p, c) {
			return true
		}
	}
	return false
}

// Unexpected states whether an error is not expected to occur. In many cases this will be due to a bug, e.g. due to a
// defensive check failing
func (p *Error) Unexpected() bool {
	if p.IsUnexpected != nil {
		return *p.IsUnexpected
	}

	return false
}

func (p *Error) SetIsRetryable(value bool) {
	if value {
		p.IsRetryable = &retryable
	} else {
		p.IsRetryable = &notRetryable
	}
}

// SetIsUnexpected can be used to explicitly mark an error as unexpected or not. In practice the vast majority of
// code should not need to use this. An example use case might be when returning a validation error that must
// mean there is a coding mistake somewhere (e.g. default statement in a switch that is never expected to be
// taken). By marking the error as unexpected there is a greater chance that an alert will be sent.
func (p *Error) SetIsUnexpected(value bool) {
	if value {
		p.IsUnexpected = &unexpected
	} else {
		p.IsUnexpected = &notUnexpected
	}
}

// LogMetadata implements the logMetadataProvider interface in the slog library which means that
// the error params will automatically be merged with the slog metadata.
// Additionally we put stack data in here for slog use.
func (p *Error) LogMetadata() map[string]string {
	return p.Params
}

// New creates a new error for you. Use this if you want to pass along a custom error code.
// Otherwise use the handy shorthand factories below
func New(code string, message string, params map[string]string) *Error {
	return errorFactory(code, message, params)
}

// NewInternalWithCause creates a new Terror from an existing error.
// The new error will always have the code `ErrInternalService`. The original
// error is attached as the `cause`, and can be tested with the `Is` function.
// You probably want to use the `Augment` func instead;
// only use this if you need to set a subcode on an error.
func NewInternalWithCause(err error, message string, params map[string]string, subCode string) *Error {
	newErr := errorFactory(errCode(ErrInternalService, subCode), message, params)
	newErr.cause = err

	switch v := err.(type) {
	case *Error:
		newErr.MessageChain = append([]string{v.Message}, v.MessageChain...)
	default:
		newErr.MessageChain = []string{err.Error()}
	}

	switch v := err.(type) {
	// If the causal error is a terror with retryability set, inherit that value.
	// Otherwise, we'll default to retryable based on the ErrInternalService code above.
	// This allows us to have an non-retryable InternalService error if the cause was not-retryable,
	// which allows the retryability of errors to propagate through the system by default, even
	// if an error handling case is missed in an upstream.
	case *Error:
		newErr.MarshalCount = v.MarshalCount
		if v.IsRetryable != nil {
			newErr.IsRetryable = v.IsRetryable
		}
	// Test if the causal error is anything else that implements the same interface and is retryable.
	case retryableError:
		r := v.Retryable()
		newErr.IsRetryable = &r
	}

	return newErr
}

type retryableError interface {
	Retryable() bool
}

// addParams returns a new error with new params merged into the original error's
func addParams(err *Error, params map[string]string) *Error {
	copiedParams := make(map[string]string, len(err.Params)+len(params))
	for k, v := range err.Params {
		copiedParams[k] = v
	}
	for k, v := range params {
		copiedParams[k] = v
	}

	return &Error{
		Code:         err.Code,
		Message:      err.Message,
		MessageChain: err.MessageChain,
		Params:       copiedParams,
		StackFrames:  err.StackFrames,
		IsRetryable:  err.IsRetryable,
		IsUnexpected: err.IsUnexpected,
		MarshalCount: err.MarshalCount,
		cause:        err.cause,
	}
}

// Matches returns whether the string returned from error.Error() contains the given param string. This means you can
// match the error on different levels e.g. dotted codes `bad_request` or `bad_request.missing_param` or even on the
// more descriptive message
// Deprecated: Please use `Is` instead. See docs for `Matches` for breaking change risks.
func (p *Error) Matches(match string) bool {
	return strings.Contains(p.Error(), match)
}

// PrefixMatches returns whether the string returned from error.Error() starts with the given param string. This means
// you can match the error on different levels e.g. dotted codes `bad_request` or `bad_request.missing_param`. Each
// dotted part can be passed as a separate argument e.g. `terr.PrefixMatches(terrors.ErrBadRequest, "missing_param")`
// is the same as `terr.PrefixMatches("bad_request.missing_param")`
// Deprecated: Please use `Is` instead.
func (p *Error) PrefixMatches(prefixParts ...string) bool {
	prefix := strings.Join(prefixParts, ".")

	return strings.HasPrefix(p.Code, prefix)
}

// Matches returns true if the error is a terror error and the string returned from error.Error() contains the given
// param string. This means you can match the error on different levels e.g. dotted codes `bad_request` or
// `bad_request.missing_param` or even on the more descriptive message
// Deprecated: Please use `Is` instead. Note that `Is` will attempt to match each error in the stack using
// `PrefixMatches`, so if you were previously matching against a part of the string returned from error.Error() that
// is _not_ the prefix, then this will be a breaking change. In this case you should update the string to match the
// prefix. If this is not possible, you can match against the entire error string explicitly, for example:
//
//	strings.Contains(err.Error(), "context deadline exceeded")
//
// But we consider this bad practice and is part of the motivation for deprecating Matches in the first place.
func Matches(err error, match string) bool {
	if terr, ok := Wrap(err, nil).(*Error); ok {
		return terr.Matches(match)
	}

	return false
}

// PrefixMatches returns true if the error is a terror and the string returned from error.Error() starts with the
// given param string. This means you can match the error on different levels e.g. dotted codes `bad_request` or
// `bad_request.missing_param`. Each dotted part can be passed as a separate argument
// e.g. `terrors.PrefixMatches(terr, terrors.ErrBadRequest, "missing_param")` is the same as
// terrors.PrefixMatches(terr, "bad_request.missing_param")`
// Deprecated: Please use `Is` instead.
func PrefixMatches(err error, prefixParts ...string) bool {
	if terr, ok := Wrap(err, nil).(*Error); ok {
		return terr.PrefixMatches(prefixParts...)
	}

	return false
}

// IsRetryable returns true if the error is a terror and whether the error was caused by an action which can be
// retried.
func IsRetryable(err error) bool {
	if r, ok := Propagate(err).(*Error); ok {
		return r.Retryable()
	}
	return false
}

// Augment adds context to an existing error.
// If the error given is not already a terror, a new terror is created.
func Augment(err error, context string, params map[string]string) error {
	if err == nil {
		return nil
	}
	switch err := err.(type) {
	case *Error:
		withMergedParams := addParams(err, params)
		// The underlying terror will already have a stack, so we don't take a new trace here.
		return &Error{
			Code:         err.Code,
			Message:      context,
			MessageChain: append([]string{err.Message}, err.MessageChain...),
			Params:       withMergedParams.Params,
			StackFrames:  stack.Stack{},
			IsRetryable:  err.IsRetryable,
			IsUnexpected: err.IsUnexpected,
			MarshalCount: err.MarshalCount,
			cause:        err,
		}
	default:
		return NewInternalWithCause(err, context, params, "")
	}
}

// Propagate an error without changing it. This is equivalent to `return err`
// if the error is already a terror. If it is not a terror, this function will
// create one, and set the given error as the cause.
// This is a drop-in replacement for `terrors.Wrap(err, nil)` which adds causal
// chain functionality.
func Propagate(err error) error {
	if err == nil {
		return nil
	}
	switch err := err.(type) {
	case *Error:
		return err
	default:
		return NewInternalWithCause(err, err.Error(), nil, "")
	}
}

// Is checks whether an error is a given code. Similarly to `errors.Is`,
// this unwinds the error stack and checks each underlying error for the code.
// If any match, this returns true.
// Note that Is only behaves differently to PrefixMatches when errors in the stack have different codes.
// For example, this is the case when errors are initialized with NewInternalWithCause, but not with Augment.
// We prefer this over using a method receiver on the terrors Error, as the function
// signature requires an error to test against, and checking against terrors would
// requite creating a new terror with the specific code.
func Is(err error, code ...string) bool {
	switch err := err.(type) {
	case *Error:
		if err.PrefixMatches(code...) {
			return true
		}
		next := err.Unwrap()
		if next == nil {
			return false
		}
		return Is(next, code...)
	default:
		return false
	}
}