The preceding discussion of how the assembly resolver determines which version of an assembly to load focuses primarily on the mechanism used by the CLR. What was not discussed is what policies a developer should use to determine when, how, and why to version an assembly. Given that the platform being described has not actually shipped at the time of this writing, it is somewhat difficult to list a set of "best practices" that are known to be valid based on hard-won experiences. However, it is reasonable to look at the known state of the CLR and extrapolate a set of guidelines.
It is important to note that assemblies are versioned as a unit. Trying to replace a subset of the files in an assembly without changing the version number will certainly lead to unpredictability. To that end, the remainder of this section looks at versioning with respect to an assembly as a whole rather than versioning individual files in an assembly.
The question of when to change version numbers is an interesting one. Obviously, if the public contract of a type changes, the type's assembly must be given a new version number. Otherwise, programs that depend on one version of the type signature will get runtime errors when a type with a different signature is loaded. This means that if you add or remove a public or protected member of a public type, you must change the version number of the type's assembly. If you change the signature of a public or protected member of a public type (e.g., adding a method parameter, changing a field's type), you also need a new assembly version number. These are absolute rules. Violating them will result in unpredictability.
The more difficult question to answer relates to modifications that do not impact the public signature of the assembly's types. For example, changes to a member that is marked as private or internal are considered nonbreaking changes, at least as far as signature matching is concerned. Because no code outside of your assembly can rely upon private or internal members, having signature mismatches occur at runtime is a nonissue because it doesn't happen. Unfortunately, signature mismatches are only the tip of the iceberg.
There is a reasonable argument to be made for changing the version number for every build of an assembly, even if no publicly visible signatures have changed. This approach is supported by the fact that even a seemingly innocuous change to a single method body may have a subtle but very real rippling effect on the behavior of programs that use the assembly. If the developer gives each build of an assembly a unique version number, code that is tested against a particular build won't be surprised at deployment time.
The argument against giving each build of an assembly a unique version number is that "safe" fixes to the code won't be picked up by programs that are not rebuilt against the new version. This argument doesn't hold water in the face of publisher policy files. Developers who use unique version numbers for every build are expected to provide publisher policy files that state which versions of their assembly are backward-compatible. By default, this gives consumers of the down-level version an automatic upgrade to the newer (and hopefully faster or less buggy) assembly. For times when the assembly's developer guesses wrong, each application can use the publisherPolicy element in its configuration file to disable the automatic upgrade, in essence running the application in "safe mode."
As discussed earlier, the CLR assembly resolver supports side-by-side installation of multiple versions of an assembly via CODEBASE hints, private probe paths, and the GAC. This allows several versions of a given assembly to peacefully coexist in the file system. However, things become somewhat unpredictable if more than one of these assemblies is actually loaded into memory at any one time, either by independent programs or by a single program. Side-by-side execution is much harder to deal with than side-by-side installation.
The primary problem with supporting multiple versions in memory at once is that, to the runtime, the types contained in those assemblies are distinct. That is, if an assembly contains a type called Customer, then when two different versions of the assembly are loaded, there are two distinct types in memory, each with its own unique identity. This has several serious downsides. For one, each type has its own copy of any static fields. If the type needed to keep track of some shared state is independent of how many versions of the type had been loaded, it could not use the obvious solution of using a static field. Rather, developers would need to rewrite the code with versioning in mind and store the shared state in a location that is not version-sensitive. One approach would be to store the shared state in some runtime-provided place such as the ASP.NET application object. Another approach would be to define a separate type that contained only the shared state as static fields. Developers could deploy this type in a separate assembly that would never be versioned, thus ensuring that only one copy of the static fields would be in memory for a given application.
Another problem related to side-by-side execution arises when versioned types are passed as method parameters. If the caller and callee of a method have differing views on which version of an assembly will be loaded, the caller will pass a parameter whose type is unknown to the callee. Developers can avoid this problem by always defining parameter types using version-invariant types for all public (cross-assembly) methods. More importantly, these shared types need to be deployed in a separate assembly that itself will not be versioned.
The metadata for an assembly has three distinguished attributes that allows the developer to specify whether multiple versions of the assembly can be loaded at the same time. If none of these attributes is present, the assembly is assumed safe for side-by-side execution in all scenarios. The nonsidebysideappdomain attribute indicates that only one version of the assembly can be loaded per AppDomain. The nonsidebysideprocess attribute indicates that only one version of the assembly can be loaded per process. Finally, the nonsidebysidemachine attribute indicates that only one version of the assembly can be loaded at a time for the entire machine. At the time of this writing, these metadata bits are ignored by the assembly resolver and loader. However,