This is a conventions RFC for formalizing the basic conventions around error handling in Rust libraries.
The high-level overview is:
For catastrophic errors, abort the process or fail the task depending on whether any recovery is possible.
For contract violations, fail the task. (Recover from programmer errors at a coarse grain.)
For obstructions to the operation, use
Result(or, less often,
Option). (Recover from obstructions at a fine grain.)
Prefer liberal function contracts, especially if reporting errors in input values may be useful to a function's caller.
Rust provides two basic strategies for dealing with errors:
Task failure, which unwinds to at least the task boundary, and by default propagates to other tasks through poisoned channels and mutexes. Task failure works well for coarse-grained error handling.
The Result type, which allows functions to signal error conditions through the value that they return. Together with a lint and the
Resultworks well for fine-grained error handling.
However, while there have been some general trends in the usage of the two handling mechanisms, we need to have formal guidelines in order to ensure consistency as we stabilize library APIs. That is the purpose of this RFC.
For the most part, the RFC proposes guidelines that are already followed today, but it tries to motivate and clarify them.
Errors fall into one of three categories:
- Catastrophic errors, e.g. out-of-memory.
- Contract violations, e.g. wrong input encoding, index out of bounds.
- Obstructions, e.g. file not found, parse error.
The basic principle of the conventions is that:
- Catastrophic errors and programming errors (bugs) can and should only be recovered at a coarse grain, i.e. a task boundary.
- Obstructions preventing an operation should be reported at a maximally fine grain -- to the immediate invoker of the operation.
An error is catastrophic if there is no meaningful way for the current task to continue after the error occurs.
Catastrophic errors are extremely rare, especially outside of
Canonical examples: out of memory, stack overflow.
For errors like stack overflow, Rust currently aborts the process, but could in principle fail the task, which (in the best case) would allow reporting and recovery from a supervisory task.
An API may define a contract that goes beyond the type checking enforced by the compiler. For example, slices support an indexing operation, with the contract that the supplied index must be in bounds.
Contracts can be complex and involve more than a single function invocation. For
RefCell type requires that
borrow_mut not be called until all
existing borrows have been relinquished.
A contract violation is always a bug, and for bugs we follow the Erlang philosophy of "let it crash": we assume that software will have bugs, and we design coarse-grained task boundaries to report, and perhaps recover, from these bugs.
One subtle aspect of these guidelines is that the contract for a function is chosen by an API designer -- and so the designer also determines what counts as a violation.
This RFC does not attempt to give hard-and-fast rules for designing contracts. However, here are some rough guidelines:
Prefer expressing contracts through static types whenever possible.
It must be possible to write code that uses the API without violating the contract.
Contracts are most justified when violations are inarguably bugs -- but this is surprisingly rare.
Consider whether the API client could benefit from the contract-checking logic. The checks may be expensive. Or there may be useful programming patterns where the client does not want to check inputs before hand, but would rather attempt the operation and then find out whether the inputs were invalid.
When a contract violation is the only kind of error a function may encounter -- i.e., there are no obstructions to its success other than "bad" inputs -- using
Optioninstead is especially warranted. Clients can then use
unwrapto assert that they have passed valid input, or re-use the error checking done by the API for their own purposes.
When in doubt, use loose contracts and instead return a
An operation is obstructed if it cannot be completed for some reason, even though the operation's contract has been satisfied. Obstructed operations may have (documented!) side effects -- they are not required to roll back after encountering an obstruction. However, they should leave the data structures in a "coherent" state (satisfying their invariants, continuing to guarantee safety, etc.).
Obstructions may involve external conditions (e.g., I/O), or they may involve aspects of the input that are not covered by the contract.
Canonical examples: file not found, parse error.
represents either a success (yielding
T) or failure (yielding
Result, a function allows its clients to discover and react to
obstructions in a fine-grained way.
Option type should not be used for "obstructed" operations; it
should only be used when a
None return value could be considered a
"successful" execution of the operation.
This is of course a somewhat subjective question, but a good litmus
test is: would a reasonable client ever ignore the result? The
Result type provides a lint that ensures the result is actually
Option does not, and this difference of behavior
can help when deciding between the two types.
Another litmus test: can the operation be understood as asking a
question (possibly with sideeffects)? Operations like
pop on a
vector can be viewed as asking for the contents of the first element,
with the side effect of removing it if it exists -- with an
An API should not provide both
failing versions of an
operation. It should provide just the
Result version, allowing clients to use
unwrap instead as needed. This is part of the general pattern of
cutting down on redundant variants by instead using method chaining.
There is one exception to this rule, however. Some APIs are strongly oriented
around failure, in the sense that their functions/methods are explicitly
intended as assertions. If there is no other way to check in advance for the
validity of invoking an operation
foo, however, the API may provide a
foo_catch variant that returns a
The main examples in
libstd that currently provide both variants are:
Channels, which are the primary point of failure propagation between tasks. As such, calling
recv()is an assertion that the other end of the channel is still alive, which will propagate failures from the other end of the channel. On the other hand, since there is no separate way to atomically test whether the other end has hung up, channels provide a
recv_optvariant that produces a
_optsuffix would be replaced by a
_catchsuffix if this RFC is accepted.
RefCell, which provides a dynamic version of the borrowing rules. Calling the
borrow()method is intended as an assertion that the cell is in a borrowable state, and will
fail!otherwise. On the other hand, there is no separate way to check the state of the
RefCell, so the module provides a
try_borrowvariant that produces a
try_prefix would be replaced by a
_catchcatch if this RFC is accepted.
(Note: it is unclear whether these APIs will continue to provide both variants.)
The main drawbacks of this proposal are:
Task failure remains somewhat of a landmine: one must be sure to document, and be aware of, all relevant function contracts in order to avoid task failure.
The choice of what to make part of a function's contract remains somewhat subjective, so these guidelines cannot be used to decisively resolve disagreements about an API's design.
The alternatives mentioned below do not suffer from these problems, but have drawbacks of their own.
As was pointed out by @SiegeLord, however, mixing what might be seen as contract violations with obstructions can make it much more difficult to write obstruction-robust code; see the linked comment for more detail.
There are numerous possible suffixes for a
_catch, as proposed above. As @kballard points out, this name connotes exception handling, which could be considered misleading. However, since it effectively prevents further unwinding, catching an exception may indeed be the right analogy.
_result, which is straightforward but not as informative/suggestive as some of the other proposed variants.
try_prefix. Also connotes exception handling, but has an unfortunately overlap with the common use of
try_for nonblocking variants (which is in play for