1.44 TensorFlow Data Versioning: GraphDefs and Checkpoints

As described in @{$version_semantics#compatibility-for-graphs-and-checkpoints$Compatibility for Graphs and Checkpoints}, TensorFlow marks each kind of data with version information in order to maintain backward compatibility. This document provides additional details about the versioning mechanism, and how to use it to safely change data formats.

Backward and partial forward compatibility

The two core artifacts exported from and imported into TensorFlow are checkpoints (serialized variable states) and GraphDefs (serialized computation graphs). Any approach to versioning these artifacts must take into account the following requirements:

• Backward compatibility to support loading GraphDefs created with older versions of TensorFlow.
• Forward compatibility to support scenarios where the producer of a GraphDef is upgraded to a newer version of TensorFlow before the consumer.
• Enable evolving TensorFlow in incompatible ways. For example, removing Ops, adding attributes, and removing attributes.

For GraphDefs, backward compatibility is enforced within a major version. This means functionality can only be removed between major versions. Forward compatibility is enforced within Patch releases (1.x.1 -> 1.x.2, for example).

In order to achieve backward and forward compatibility as well as know when to enforce changes in formats, the serialized representations of graphs and variable state need to have metadata that describes when they were produced. The sections below detail the TensorFlow implementation and guidelines for evolving GraphDef versions.

Independent data version schemes

There are data versions for GraphDefs and checkpoints. Both data formats evolve at different rates, and also at different speeds than the version of TensorFlow. Both versioning systems are defined in core/public/version.h. Whenever a new version is added a note is added to the header detailing what changed and the date.

Data, producers, and consumers

This section discusses version information for data, binaries that produce data (producers), and binaries that consume data (consumers):

• Producer binaries have a version (producer) and a minimum consumer version that they are compatible with (min_consumer).
• Consumer binaries have a version (consumer) and a minimum producer version that they are compatible with (min_producer).
• Each piece of versioned data has a VersionDef versions field which records the producer that made the data, the min_consumer that it is compatible with, and a list of bad_consumers versions that are disallowed.

By default, when a producer makes some data, the data inherits the producer's producer and min_consumer versions. bad_consumers can be set if specific consumer versions are known to contain bugs and must be avoided. A consumer can accept a piece of data if

• consumer >= data's min_consumer
• data's producer >= consumer's min_producer
• consumer not in data's bad_consumers

Since both producers and consumers come from the same TensorFlow code base, core/public/version.h contains a main binary version which is treated as either producer or consumer depending on context and both min_consumer and min_producer (needed by producers and consumers, respectively). Specifically,

• For GraphDef versions, we have TF_GRAPH_DEF_VERSION, TF_GRAPH_DEF_VERSION_MIN_CONSUMER, and TF_GRAPH_DEF_VERSION_MIN_PRODUCER.
• For checkpoint versions, we have TF_CHECKPOINT_VERSION, TF_CHECKPOINT_VERSION_MIN_CONSUMER, and TF_CHECKPOINT_VERSION_MIN_PRODUCER.

Evolving GraphDef versions

This section presents examples of using this versioning mechanism to make changes to the GraphDef format.

Adding a new Op:

1. Add the new Op to both consumers and producers at the same time, and do not change any GraphDef versions. This type of change is automatically backward compatible, and does not impact forward compatibility plan since existing producer scripts will not suddenly use the new functionality.

Adding a new Op and switching existing Python wrappers to use it:

1. Implement new consumer functionality and increment the binary version.
2. If it is possible to make the wrappers use the new functionality only in cases that did not work before, the wrappers can be updated now.
3. Change Python wrappers to use the new functionality. Do not increment min_consumer, since models which do not use this Op should not break.

Removing an Op or restricting the functionality of an Op:

1. Fix all producer scripts (not TensorFlow itself) to not use the banned Op or functionality.
2. Increment the binary version and implement new consumer functionality that bans the removed Op or functionality for GraphDefs at the new version and above. If possible, make TensorFlow stop producing GraphDefs with the banned functionality. This can be done with REGISTER_OP(...).Deprecated(deprecated_at_version, message).
3. Wait for a major release for backward compatibility purposes.
4. Increase min_producer to the GraphDef version from (2) and remove the functionality entirely.

Changing the functionality of an Op:

1. Add a new similar Op named SomethingV2 or similar and go through the process of adding it and switching existing Python wrappers to use it (may take 3 weeks if forward compatibility is desired).
2. Remove the old Op (Can only take place with a major version change due to backward compatibility).
3. Increase min_consumer to rule out consumers with the old Op, add back the old Op as an alias for SomethingV2, and go through the process to switch existing Python wrappers to use it.
4. Go through the process to remove SomethingV2.

Banning a single consumer version that cannot run safely:

1. Bump the binary version and add the bad version to bad_consumers for all new GraphDefs. If possible, add to bad_consumers only for GraphDefs which contain a certain Op or similar.
2. If existing consumers have the bad version, push them out as soon as possible.