Permalink
Fetching contributors…
Cannot retrieve contributors at this time
408 lines (306 sloc) 16.8 KB

Trillian: General Transparency

Build Status Go Report Card GoDoc Slack Status

Overview

Trillian is an implementation of the concepts described in the Verifiable Data Structures white paper, which in turn is an extension and generalisation of the ideas which underpin Certificate Transparency.

Trillian implements a Merkle tree whose contents are served from a data storage layer, to allow scalability to extremely large trees. On top of this Merkle tree, Trillian provides two modes:

  • An append-only Log mode, analogous to the original Certificate Transparency logs. In this mode, the Merkle tree is effectively filled up from the left, giving a dense Merkle tree.
  • A Map mode that allows transparent storage of arbitrary key:value pairs. In this mode, the key's hash is used to designate a particular leaf of a deep Merkle tree, giving a sparse Merkle tree. (A Trillian Map is an unordered map; it does not allow enumeration of the Map's keys.)

Note that Trillian requires particular applications to provide their own personalities on top of the core transparent data store functionality; example code for a certificate transparency log and for a log-derived map are included to help with this.

The code for the CT personality has now been moved to a separate repository and can be obtained from certificate-transparency-go.

Support

Using the Code

WARNING: The Trillian codebase is still under development but is now being used in production by several organizations. We will try to avoid any further incompatible code and schema changes but cannot guarantee that they will never be necessary.

To build and test Trillian you need:

  • Go 1.9 or later.

To run many of the tests (and production deployment) you need:

Use the standard Go tools to install other dependencies.

go get github.com/google/trillian
cd $GOPATH/src/github.com/google/trillian
go get -t -u -v ./...

To build and run tests, use:

go test ./...

Note that go seems to sometimes fail to fetch or update all dependencies (as of v1.10.2), so you may need to manually fetch missing ones, or update all Go source with:

go get -u -v all

The repository also includes multi-process integration tests, described in the Integration Tests section below.

MySQL Setup

To run Trillian's integration tests you need to have an instance of MySQL running and configured to:

  • listen on the standard MySQL port 3306 (so mysql --host=127.0.0.1 --port=3306 connects OK)
  • not require a password for the root user

You can then set up the expected tables in a test database like so:

./scripts/resetdb.sh
Warning: about to destroy and reset database 'test'
Are you sure? y
> Resetting DB...
> Reset Complete

If you are working with the Trillian Map, you will probably need to increase the MySQL maximum connection count:

% mysql -u root
MySQL> SET GLOBAL max_connections = 1000;

Integration Tests

Trillian includes an integration test suite to confirm basic end-to-end functionality, which can be run with:

./integration/integration_test.sh

This runs two multi-process tests:

  • A test that starts a Trillian server in Map mode, sets various key:value pairs and checks they can be retrieved.
  • A test that starts a Trillian server in Log mode, together with a signer, logs many leaves, and checks they are integrated correctly.

Working on the Code

Developers who want to make changes to the Trillian codebase need some additional dependencies and tools, described in the following sections. The Travis configuration for the codebase is also useful reference for the required tools and scripts, as it may be more up-to-date than this document.

Rebuilding Generated Code

Some of the Trillian Go code is autogenerated from other files:

  • gRPC message structures are originally provided as protocol buffer message definitions.
  • Some unit tests use mock implementations of interfaces; these are created from the real implementations by GoMock.
  • Some enums have string-conversion methods (satisfying the fmt.Stringer interface) created using the stringer tool (go get golang.org/x/tools/cmd/stringer).

Re-generating mock or protobuffer files is only needed if you're changing the original files; if you do, you'll need to install the prerequisites:

and run the following:

go generate -x ./...  # hunts for //go:generate comments and runs them

Updating Vendor Code

The Trillian codebase includes a couple of external projects under the vendor/ subdirectory, to ensure that builds use a fixed version (typically because the upstream repository does not guarantee back-compatibility between the tip master branch and the current stable release). These external codebases are included as Git subtrees.

To update the code in one of these subtrees, perform steps like:

# Add master repo for upstream code as a Git remote.
git remote add vendor-xyzzy https://github.com/orgname/xyzzy
# Pull the updated code for the desired version tag from the remote, dropping history.
# Trailing / in prefix is needed.
git subtree pull --squash --prefix=vendor/github.com/orgname/xyzzy/ vendor-xyzzy vX.Y.Z

If new vendor/ subtree is required, perform steps similar to:

# Add master repo for upstream code as a Git remote.
git remote add vendor-xyzzy https://github.com/orgname/xyzzy
# Pull the desired version of the code in, dropping history.
# Trailing / in --prefix is needed.
git subtree add --squash --prefix=vendor/github.com/orgname/xyzzy/ vendor-xyzzy vX.Y.Z

Running Codebase Checks

The scripts/presubmit.sh script runs various tools and tests over the codebase.

# Install gometalinter and all linters
go get -u github.com/alecthomas/gometalinter
gometalinter --install

# Run code generation, build, test and linters
./scripts/presubmit.sh

# Or just run the linters alone:
gometalinter --config=gometalinter.json ./...

Design

Design Overview

Trillian is primarily implemented as a gRPC service; this service receives get/set requests over gRPC and retrieves the corresponding Merkle tree data from a separate storage layer (currently using MySQL), ensuring that the cryptographic properties of the tree are preserved along the way.

The Trillian service is multi-tenanted – a single Trillian installation can support multiple Merkle trees in parallel, distinguished by their TreeId – and operates in one of two modes:

  • Log mode: an append-only collection of items.
  • Map mode: a collection of key:value pairs.

In either case, Trillian's key transparency property is that cryptographic proofs of inclusion/consistency are available for data items added to the service.

Personalities

The Trillian service expects to be paired with additional code that is specific to the particular application of the transparent store; this is known as a personality.

The primary purpose of a personality is to implement admission criteria for the store, so that only particular types of data are added to the store. For example, a certificate transparency log only accepts data items that are valid certificates; a "CT Log" personality would police this, so that the Trillian service can process all incoming data blindly.

A personality may also perform canonicalization on incoming data, to convert equivalent formulations of the same underlying data to a single canonical format, avoiding needless duplication. (For example, keys in JSON dictionaries could be sorted, or Unicode string data could be normalised.)

The per-application personality is also responsible for providing an externally-visible interface, typically over HTTP[S].

Note that a personality may need to implement its own data store, separate from Trillian. In particular, if the personality does not completely trust Trillian, it needs to store the various things that Trillian signs in order to be able to detect problems (and so the personality effectively also acts as a monitor for Trillian).

Map Mode

Trillian in Map mode can be thought of as providing a key:value store, together with cryptographic transparency guarantees for that data.

When running in Map mode, Trillian provides a straightforward gRPC API with the following available operations:

  • GetSignedMapRoot returns information about the current root of the Merkle tree representing the Map, including a revision (see below), hash value, timestamp and signature.
  • GetLeaves returns leaf information for a specified set of key values, optionally as of a particular revision. The returned leaf information also includes inclusion proof data.
  • SetLeaves requests inclusion of specified key:value pairs into the Map; these will appear as the next revision of the Map.

(Documentation may be out-of-date; please check the protocol buffer message definitions for the definitive current map API.)

Each SetLeaves request includes a batch of updates to the Map; once all of these updates have been applied, the Map has a new revision, with a new tree head for that revision. To allow historical queries, the API allows queries of the Map as of a particular revision.

TODO: add description of per-personality Mappers

TODO: add description of distribution: how many instances run, how distributed, how synchronized (master election), mention use of transactions as a fallback (in case of errors in master election).

Map components

Log Mode

When running in Log mode, Trillian provides a gRPC API whose operations are similar to those available for Certificate Transparency logs (cf. RFC 6962). These include:

  • GetLatestSignedLogRoot returns information about the current root of the Merkle tree for the log, including the tree size, hash value, timestamp and signature.
  • GetLeavesByHash, GetLeavesByIndex and GetLeavesByRange return leaf information for particular leaves, specified either by their hash value or index in the log.
  • QueueLeaves requests inclusion of specified items into the log.
  • GetInclusionProof, GetInclusionProofByHash and GetConsistencyProof return inclusion and consistency proof data.

In Log mode, Trillian includes an additional Signer component; this component periodically processes pending queued items and adds them to the Merkle tree, creating a new signed tree head as a result.

Log components

TODO: add description of distribution: how many instances run, how distributed etc.

Logged Map

As it currently stands, it is not possible to reliably monitor or audit a Trillian Map instance; key:value pairs can be modified and subsequently reset without anyone noticing.

A future plan to deal with this is to create a Logged Map, which combines a Trillian Map with a Trillian Log so that all published revisions of the Map have their signed tree head data appended to the corresponding Map.

Use Cases

Certificate Transparency Log

The most obvious application for Trillian in Log mode is to provide a certificate transparency (RFC 6962) Log. To do this, the CT Log personality needs to include all of the certificate-specific processing – in particular, checking that an item that has been suggested for inclusion is indeed a valid certificate that chains to an accepted root.

Verifiable Log-Derived Map

One useful application for Trillian in Map mode is to provide a verifiable log-derived map (VLDM), as described in the Verifiable Data Structures white paper (which uses the term 'log-backed map'). To do this, a VLDM personality would monitor the additions of entries to a Log, potentially external, and would write some kind of corresponding key:value data to a Trillian Map.

Clients of the VLDM are then able to verify that the entries in the Map they are shown are also seen by anyone auditing the Log for correct operation, which in turn allows the client to trust the key/value pairs returned by the Map.

A concrete example of this might be a VLDM that monitors a certificate transparency Log and builds a corresponding Map from domain names to the set of certificates associated with that domain.

The following table summarizes properties of data structures laid in the Verifiable Data Structures white paper. “Efficiently” means that a client can and should perform this validation themselves. “Full audit” means that to validate correctly, a client would need to download the entire dataset, and is something that in practice we expect a small number of dedicated auditors to perform, rather than being done by each client.

Verifiable Log Verifiable Map Verifiable Log-Derived Map
Prove inclusion of value Yes, efficiently Yes, efficiently Yes, efficiently
Prove non-inclusion of value Impractical Yes, efficiently Yes, efficiently
Retrieve provable value for key Impractical Yes, efficiently Yes, efficiently
Retrieve provable current value for key Impractical No Yes, efficiently
Prove append-only Yes, efficiently No Yes, efficiently [1].
Enumerate all entries Yes, by full audit Yes, by full audit Yes, by full audit
Prove correct operation Yes, efficiently No Yes, by full audit
Enable detection of split-view Yes, efficiently Yes, efficiently Yes, efficiently
  • [1] -- although full audit is required to verify complete correct operation