wiki:Commentary/Packages/Concepts

Version 1 (modified by ezyang, 4 years ago) (diff)

--

Package keys, installed package IDs, ABI hashes, package names and versions, Nix-style hashes, ... there's so many different identifiers, what do they all mean? I think the biggest source of confusion (for myself included) is keeping straight not only what these terms mean, but also what people want them to mean in the future, and what we actually care about. So I want to help clarify this a bit, by clearly separating the motivation from the mechanism.

The content here overlaps with wiki:Commentary/Packages but is looking at the latest iteration of the multi-instances and Backpack work.

Motivations

They are a number of different ways we can identify a package, driven by different MOTIVATIONS:

[SYMBOL]
What symbol names should we put in the binary? (e.g., the "foozm0zi1" in "foozm0zi1_A_DZCF_closure")
  • It must be unique enough that for all libraries we would like to be able to link together, there should not be conflicts.
  • HOWEVER, it must be stable enough that if we make a minor source code change, we don't have to gratuitously recompile every dependency.
[ABI]
When can I swap out one compiled package with another WITHOUT recompiling, i.e. what is the ABI of the package? Equal ABIs implies equal symbols.
[SOURCE]
What is the specific stream of bytes that represents the source package? In Hackage, this constitutes the unit of distribution, i.e. how do you identify the source tarball a package maintainer uploads to Hackage.
  • On Hackage, a package name plus version uniquely identifies an sdist. This is enforced by community standards; in a local development environment, this may not hold since devs will edit code without updating the version number.
  • Alternately, a cryptographic hash of the source code uniquely identifies the stream of bytes. This is enforced by math.
[NIX]
What is the specific stream of bytes that represents the source package and the source of all of its transitive dependencies? Put alternately, what is the full stream of bytes that, in isolation, would be sufficient to reproduce a build (assuming a deterministic compiler.)
[TYPES]
When are two types the same? If there are from differing packages, they are obviously different; if they are from the same package, they might still be different if the dependencies were different in each case.
  • Types show up in error message, so this is a USER VISIBLE notion. Many people have (cogently) argued that this should be AS SIMPLE as possible, because there's nothing worse than being told that Data.ByteString.ByteString is not equal to Data.ByteString.ByteString (because they were from different packages.)

Mechanisms

Today, we have a lot of different MECHANISMS for identifying these:

Package Name
Something like "lens"
Package Version
Something like "0.1.2"
Package ID
Package name plus version. By today's Hackage conventions, this can identify a [SOURCE]. Pre-GHC 7.10, it also identified [SYMBOL] and [TYPES].
Installed Package ID
Package name, package version, and the output of ghc --abi-hash. This identifies [ABI].
Package Key (new in 7.10)
Hash of package name, package version, the package keys of all direct dependencies the package was compiled against, and a a mapping of how Backpack holes in the package were instantiated. This now identifies [SYMBOL] and [TYPES]. It also identifies [NIX] if you assume that package IDs adequately identify a [SOURCE].

Furthermore, there have been some proposals for how things should further change in the future. Here are the few that I am aware of:

Installed Package ID (WITHOUT ABI)
In our conversations last summer, Duncan reported to me that he wanted to change Installed Package ID so that it was a hash of the source code and other relevant bits, serving the role of [NIX] hashes. I recorded this in this Cabal bug https://github.com/haskell/cabal/issues/2199
Package Key (WITHOUT KEYS OF INCLUDES)
While I was working on Backpack I discovered that our current definition of package key is circular (and thus undefined) in some use-cases which we might eventually want to support. The refinement I currently have in my working Git copy is that I have removed from the hash the package keys of the direct dependencies of the package. If you apply this change, package keys NO LONGER identify a [NIX].
Package KEY (WITH SOURCE HASH)
In this 2015 Haskell GSoC proposal https://gist.github.com/fugyk/37510958b52589737274 Vishal Agrawal proposes adding a hash of the source tarball to the package key. This would make it implement [NIX] in the strong sense. However, this proposal would make package key unsuitable for [SOURCE]

Features

There are a number of enhancements proposed for how Cabal handles packages, which have often been conflated together. I want to clearly separate them out here:

Non-destructive installs
If I have package foo-0.2 compiled against bar-0.1, and a different build compiled against bar-0.2, I should be able to put them in the same installed package database. THIS IS HIGH PRIORITY.
Views
If I have package foo compiled against bar-0.1, and baz compiled against bar-0.2, these two packages aren't usable together (modulo private dependencies, see below). Views are a UI paradigm making it easier for users to work in a universe where foo is available, or a universe where baz is available, but not both simultaneously. Cabal sandboxes are views but without a shared installed package database. This is lower priority, because if you use cabal-install to get a coherent dependency set, you'll never see both foo and baz at the same time; the primary benefit of this is to assist with direct use of GHC/GHCi, however, it is generally believed that non-destructive installs will make it difficult to use GHC/GHCi by itself.
Private dependencies
If I have a package foo-0.2 which depends on a library bar-0.1, but not in any externally visible way, it should be allowed for a client to separately use bar-0.2. This is LOWEST priority; amusingly, in 7.10, this is already supported by GHC, but not by Cabal.