Design Document: Implementing Beautiful Error Messages (Loading Phase)

Design documents are not descriptions of the current functionality of Bazel. Always go to the documentation for current information.

Status: unimplemented


Design document published: 26 May 2016

Related: "Beautiful error messages"


This is a followup to the document "Beautiful error messages".

The purpose of this document is to outline a design for some plumbing that will allow the sort of errors described in that document to be emitted by Blaze for loading time BUILD file errors.

Review: What needs to be done

In the "Beautiful error messages" document, four characteristics of error messages are enumerated:

  1. Context: The erroneous text in the BUILD file in question is shown, with a caret pointing at the exact expression in question.

  2. Colors: Everyone loves colors.

  3. Suggestions: A guess of how the error should be fixed is shown.

  4. Links: Documentation is referenced directly in error messages.

This document covers (1) and (3).

Current Error Infrastructure

In the loading phase, Bazel parses BUILD files into a tree of ASTNode instances, with a BuildFileAST at the root and Statement instances at the leaves. Each statement implements doExec, which can throw an EvalException. That exception is translated to the error printed to the terminal.

An EvalException encapsulates the information given in an error. Fleshing out the contents of this exception type is a good starting point for implementing new error features.

Implementation: Context

An EvalException contains a reference to a Location, which encapsulates the line/character information currently specified in a Bazel error.

Generally, an ASTNode will have a location instance that is populated by the parser as it moves from token to token (example). Right now, since the location only contains pure syntactic information, the parser calls into the lexer to create the location instance. While the parser perhaps could also encode AST information into the Location, that shouldn’t be necessary to provide "context" in the form of a printed line and a carat. It seems that the lexer maintains BUILD file info as a buffer, and should be able to parameterize Location instances with the actual contents of the line in question. If this is true, then implementing (1) above shouldn’t involve much more than fleshing out LocationInfo.

Implementation: Suggestions

While data tracked by the lexer should be sufficient to encode an offending line from a BUILD file into an error, providing suggestions will probably require semantic information only retrievable from the parsed AST. Furthermore, we require a mechanism in EvalException to perform computation on AST data in order to generate suggestions.

It seems an unlikely solution to encode AST information in Location instances, since those instances are produced by the parser before the AST, or even the AST node in question, is necessarily complete.

Instead, here are a couple proposals:

  1. An abstract subclass of EvalException (e.g. ContextualEvalException) that knows how to create an error message with suggestions given unimplemented suggestion generation logic.

  2. A further group of exceptions that are parameterized with a particular sort of ASTNode and know how to generate suggestions. As an example, this exception type could report typos in rule names.

public class IdentifierEvalException extends ContextualEvalException {
 public static final Map<String, String> TYPOS = Map.of(
    "cclibrary", "cc_library",
  public IdentifierEvalException(Identifier identifier, Location loc) {...}
  protected Suggestion generateSuggestion() {
    if (TYPOS.keySet().contains(identifier.getName()) {
      return TypoSuggestion(identifier, TYPOS.get(identifier.getName()));

Once the plumbing around (1) is in place, we can add subclasses at our leisure to provide suggestions.

Furthermore, a ContextualEvalException can be made to have enough information to not only provide suggestions, but also to return context-aware error messages. Consider this example, from the "Beautiful error messages" document:

my_obj = select({
  ":something": [1],
  "other": [2],

t = [x for x in my_obj]
ERROR: /test/BUILD:6:5: type 'select' is not iterable.

ERROR: /test/BUILD:6:16: my_obj of type 'select' is not iterable.
You can iterate only on strings, lists, tuples or dicts.
t = [x for x in my_obj]
Related documentation: http://documentation#select

We can imagine a NotIterableEvalException that knows not only about the type select, but is also parameterized with the erroneous expression my_objc.

Problem: Serialization

The above proposal hinges on the ability to store a file pointer in the Location object, to be dereferenced at error-time to obtain the entire BUILD file. This opens the door to some issues:

  1. A Location instance is stored for every node in the parse tree of every BUILD file. Even a file pointer in each Location may have substantial memory/speed impact.

    This impact is easily measurable and likely tolerable in order to achieve better error messages. However, it is clear that storing anything much larger than a file pointer (like a fragment of the file, or the file itself) in each Location would be untenable.

  2. The Location object is serialized, since the AST is part of a SkyValue. This pointer, then, must be serializable.

    This in particular is troubling because, unlike the java heap in a local Bazel execution, the SkyValue containing the AST does not necessarily have the BUILD file. However, it is not clear to me that this means the file must be copied into each Location object. The AST presumably resides in a single SkyValue - one copy of the file in that SkyValue, with a pointer to that file in the Location, would be sufficient, it seems.

    The nature of the SkyValue that contains that AST must be determined, thinking about if and how to embed the BUILD file into that SkyValue, and strategizing about a good serialization for a Location object that contains a file pointer.