Actually, that last revision can be simpler -- we don't need to parse strings at all, as simply entering 1e1000 as a float literal in Python will be evaluated as infinity.

In Python, avoid relying on float('inf') and float('nan') as these don't work on Windows with Python pre-2.6.

C++ and Java.  See CHANGES.txt for more details.

for details.

All Languages
* Repeated fields of primitive types (types other that string, group, and
  nested messages) may now use the option [packed = true] to get a more
  efficient encoding.  In the new encoding, the entire list is written
  as a single byte blob using the "length-delimited" wire type.  Within
  this blob, the individual values are encoded the same way they would
  be normally except without a tag before each value (thus, they are
  tightly "packed").

C++
* UnknownFieldSet now supports STL-like iteration.
* Message interface has method ParseFromBoundedZeroCopyStream() which parses
  a limited number of bytes from an input stream rather than parsing until
  EOF.

Java
* Fixed bug where Message.mergeFrom(Message) failed to merge extensions.
* Message interface has new method toBuilder() which is equivalent to
  newBuilderForType().mergeFrom(this).
* All enums now implement the ProtocolMessageEnum interface.
* Setting a field to null now throws NullPointerException.
* Fixed tendency for TextFormat's parsing to overflow the stack when
  parsing large string values.  The underlying problem is with Java's
  regex implementation (which unfortunately uses recursive backtracking
  rather than building an NFA).  Worked around by making use of possesive
  quantifiers.

Python
* Updated RPC interfaces to allow for blocking operation.  A client may
  now pass None for a callback when making an RPC, in which case the
  call will block until the response is received, and the response
  object will be returned directly to the caller.  This interface change
  cannot be used in practice until RPC implementations are updated to
  implement it.

protoc
* Enum values may now have custom options, using syntax similar to field
  options.
* Fixed bug where .proto files which use custom options but don't actually
  define them (i.e. they import another .proto file defining the options)
  had to explicitly import descriptor.proto.
* Adjacent string literals in .proto files will now be concatenated, like in
  C.

C++
* Generated message classes now have a Swap() method which efficiently swaps
  the contents of two objects.
* All message classes now have a SpaceUsed() method which returns an estimate
  of the number of bytes of allocated memory currently owned by the object.
  This is particularly useful when you are reusing a single message object
  to improve performance but want to make sure it doesn't bloat up too large.
* New method Message::SerializeAsString() returns a string containing the
  serialized data.  May be more convenient than calling
  SerializeToString(string*).
* In debug mode, log error messages when string-type fields are found to
  contain bytes that are not valid UTF-8.
* Fixed bug where a message with multiple extension ranges couldn't parse
  extensions.
* Fixed bug where MergeFrom(const Message&) didn't do anything if invoked on
  a message that contained no fields (but possibly contained extensions).
* Fixed ShortDebugString() to not be O(n^2).  Durr.
* Fixed crash in TextFormat parsing if the first token in the input caused a
  tokenization error.

Java
* New overload of mergeFrom() which parses a slice of a byte array instead
  of the whole thing.
* New method ByteString.asReadOnlyByteBuffer() does what it sounds like.
* Improved performance of isInitialized() when optimizing for code size.

Python
* Corrected ListFields() signature in Message base class to match what
  subclasses actually implement.
* Some minor refactoring.

  it was valid or not.
* Moved UTF-8 coding directive to the right place in reflection_test.py.

  General
  * License changed from Apache 2.0 to New BSD.
  * It is now possible to define custom "options", which are basically
    annotations which may be placed on definitions in a .proto file.
    For example, you might define a field option called "foo" like so:
      import "google/protobuf/descriptor.proto"
      extend google.protobuf.FieldOptions {
        optional string foo = 12345;
      }
    Then you annotate a field using the "foo" option:
      message MyMessage {
        optional int32 some_field = 1 [(foo) = "bar"]
      }
    The value of this option is then visible via the message's
    Descriptor:
      const FieldDescriptor* field =
        MyMessage::descriptor()->FindFieldByName("some_field");
      assert(field->options().GetExtension(foo) == "bar");
    This feature has been implemented and tested in C++ and Java.
    Other languages may or may not need to do extra work to support
    custom options, depending on how they construct descriptors.

  C++
  * Fixed some GCC warnings that only occur when using -pedantic.
  * Improved static initialization code, making ordering more
    predictable among other things.
  * TextFormat will no longer accept messages which contain multiple
    instances of a singular field.  Previously, the latter instance
    would overwrite the former.
  * Now works on systems that don't have hash_map.

  Python
  * Strings now use the "unicode" type rather than the "str" type.
    String fields may still be assigned ASCII "str" values; they will
    automatically be converted.
  * Adding a property to an object representing a repeated field now
    raises an exception.  For example:
      # No longer works (and never should have).
      message.some_repeated_field.foo = 1

protoc
- New flags --encode and --decode can be used to convert between protobuf text
  format and binary format from the command-line.
- New flag --descriptor_set_out can be used to write FileDescriptorProtos for
  all parsed files directly into a single output file.  This is particularly
  useful if you wish to parse .proto files from programs written in languages
  other than C++: just run protoc as a background process and have it output
  a FileDescriptorList, then parse that natively.

C++
- Reflection objects are now per-class rather than per-instance.  To make this
  possible, the Reflection interface had to be changed such that all methods
  take the Message instance as a parameter.  This change improves performance
  significantly in memory-bandwidth-limited use cases, since it makes the
  message objects smaller.  Note that source-incompatible interface changes
  like this will not be made again after the library leaves beta.

Python
- MergeFrom(message) and CopyFrom(message) are now implemented.
- SerializeToString() raises an exception if the message is missing required
  fields.
- Code organization improvements.
- Fixed doc comments for RpcController and RpcChannel, which had somehow been
  swapped.

Protoc (parser)
- Improved error message when an enum value's name conflicts with another
  symbol defined in the enum type's scope, e.g. if two enum types declared
  in the same scope have values with the same name.  This is disallowed for
  compatibility with C++, but this wasn't clear from the error.
C++
- Restored the set_foo(const char*) accessor for "bytes" type because some
  code inside Google depends on it.  However, set_foo(const char*, int) is
  still there (and actually is changed to take const void*).
- Fixed TokenizerTest when compiling with -DNDEBUG on Linux.
- Other irrelevant tweaks.
Java
- Fixed UnknownFieldSet's parsing of varints larger than 32 bits.
- Fixed TextFormat's parsing of "inf" and "nan".
- Fixed TextFormat's parsing of comments.
Python
- Fixed text_format_test on Windows where floating-point exponents sometimes
  contain extra zeros.

Patch from Johan Euphrosine <proppy@aminche.com>