TensorFlow Data Versioning: GraphDefs and Checkpoints

As described in Compatibility for Graphs and Checkpoints, TensorFlow marks each kind of data with version information in order to maintain backwards compatibility even across major releases in some cases.

This document describes the versioning mechanism in more detail, and explains how to use it to change data formats safely.

Goals: backwards and partial forwards compatibility

Consider the case of TensorFlow graphs serialized via the GraphDef protobuf. We have a number of competing constraints:

  • We would like to be able to evolve TensorFlow in eventually incompatible ways: removing ops, adding or removing attrs, etc.
  • GraphDefs produced by TensorFlow may live for months after they are generated, so we want backwards compatibility: new versions of TensorFlow should be able to read old data.
  • Sometimes a producer of a GraphDef is upgraded to a new version of TensorFlow before the consumer of that data is updated, so we would like forwards compatibility: new versions of TensorFlow should generate GraphDefs readable by older versions of TensorFlow. Unfortunately, forwards compatibility is much more intrusive than backwards compatibility, so we support it only in limited situations within Google and across patch releases for open source.

For GraphDefs, we support backwards compatibility for 6 months and forwards compatibility for 3 weeks in limited situations. For backwards compatibility, this means that we can only remove functionality 6 months after we stop producing data using that functionality. Similarly, in the limited situations where we support forwards compatibility, we can add functionality only 3 weeks after TensorFlow can consume data using that functionality.

In order to implement these semantics, we need to know when data is produced so that we can know when to enforce changes in formats. The versioning system described below achieves that goal in a manner that supports both backwards and forwards compatibility (when they apply).

For checkpoints, we have no plans to make either backwards or forwards incompatible changes, but still attach versions to checkpoints in case we ever do have to make a change.

Each type of data has separate version scheme

Since different data formats evolve at different rates, we have a separate integer versioning scheme for each kind of data, and these schemes are separate from the overall version of TensorFlow.

For now, there are data versions for GraphDefs (serialized computation graphs) and checkpoints (serialized variable state). Both versioning schemes are defined in core/public/version.h. Whenever a new version is added, a note should be made in that header recording what changed and when.

Data, producers, and consumers

In the discussion below, we consider version information for data, binaries that produce that data (producers), and binaries that consume that 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,


Evolving GraphDef versions

We now discuss examples of using this versioning mechanism to make various changes to the GraphDef format. Our goal is to be backwards compatible for six months, which means that data produced by TensorFlow at time T must be consumable by TensorFlow at time T + 6 months. If forwards compatibility is desired, the data must be consumable at time T - 3 weeks.

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 backwards compatible, and is outside our forwards 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. If forwards compatibility is necessary, wait 3 weeks.
  4. 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 6 months for backwards 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 forwards compatibility is desired).
  2. Remove the old op (takes 6 months due to backwards 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 (may take 3 weeks).
  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.