bdk_core: a new architecture for the Bitcoin Dev Kit

By Lloyd Fournier on 5/10/2022 - Tags: Architecture


The Bitcoin Devkit (BDK) lets you do a lot of useful things through convenient high level abstractions. When these abstractions map nicely to your application components your life will be easy but when they don't it will be frustrating. My rather ambitious plan is to start developing a new bdk_core library that exposes all the useful mechanisms that BDK has inside it without them being tied to any particular usage policy and with very minimal dependencies.

The bdk_core idea is still "in the lab". We're not sure yet whether bdk_core will just be what's left of bdk once we spin off all the components that have extra dependencies into their own crates and refine it a bit. In that case bdk_core will just be called bdk v1.0.0 or something. Or it might be that bdk lives on with its current APIs and uses stuff bdk_core to implement it internally.

# The separation of policy and mechanism

My guiding principle for bdk_core is going to be the separation of policy and mechanism. This is what I mean by these terms:

  • mechanism: How you do a particular thing. Mechanism code is functional and doesn't change much.
  • policy: What you want to do. Policy code composes mechanisms to achieve something in an application.

Here's a nice passage about why the designers of the X window system (opens new window) applied this principle. X has been around since 1984 and doesn't look like it's going anywhere so it probably has a lot to teach us. From The Art of UNIX Programming (opens new window):

...we observed that the designers of X made a basic decision to implement “mechanism, not policy”—to make X a generic graphics engine and leave decisions about user-interface style to toolkits and other levels of the system. We justified this by pointing out that policy and mechanism tend to mutate on different timescales, with policy changing much faster than mechanism. Fashions in the look and feel of GUI toolkits may come and go, but raster operations and compositing are forever.

Thus, hardwiring policy and mechanism together has two bad effects: It makes policy rigid and harder to change in response to user requirements, and it means that trying to change policy has a strong tendency to destabilize the mechanisms.

On the other hand, by separating the two we make it possible to experiment with new policy without breaking mechanisms. We also make it much easier to write good tests for the mechanism (policy, because it ages so quickly, often does not justify the investment).

This design rule has wide application outside the GUI context. In general, it implies that we should look for ways to separate interfaces from engines.

You'll notice we have a similar situation in Bitcoin engineering. We have mechanism code like signing algorithms, key derivation, transaction construction logic, etc., that don't change much. But how these compose together in applications changes quickly over time and between applications.

The main culprit of policy and mechanism conflation in bdk is the main Wallet (opens new window) type. Wallets do all of the following:

  1. Store one or two descriptors (external and optional internal).
  2. Keep track of which addresses you've given out so you only give out fresh ones from each descriptor.
  3. Keep a list of transactions associated with the addresses in the wallet.
  4. Given a source of blockchain data it can update its internal list of transactions.
  5. Given some parameters it can build a PSBT from transaction outputs.
  6. Given a PSBT it can sign it with its Signers (opens new window).

All of that is very useful but it is bound together with the particular policies and opinions of Wallet. If Wallet's policy is not your policy it's going to be tricky to get it to do what you want. Here are some examples:

  1. In order to control how the Wallet will select coins for a transaction internally you have to pass in something implementing the CoinSelectionAlgorithm (opens new window) trait. A coin selection algorithm is clearly mechanism code but the policy of Wallet restricts that mechanism's interface. We have very old issues (opens new window) related to what the interface of this trait should be and we don't have a clear way forward. In bdk_core I want to purely provide the coin selection mechanisms for figuring out whether you need to select more UTXOs or whether you need a change output etc. How you use that mechanism will be up to you.
  2. Another trait that has a similar structure is the Signer (opens new window) trait. You have to pass in signers so your wallet can sign PSBTs but you have little control over how the wallet chooses which signers to use in any given situation. Right now the wallet will just iterate through all the signers and ask them to sign. This is not always appropriate. In bdk_core I hope we can just provide ways of populating the correct PSBT fields with signatures.

# A syncing mechansim without the policy

The rest of this post goes into detail about how to expose useful mechanisms for syncing without imposing a policy either on how blockchain data is fetched or how it is stored.

# The problems with the policy

Syncing in bdk is the place where the design of Wallet is most restrictive. The WalletSync (opens new window) trait forces you to sync all addresses in a wallet in one big batch. But this is not always what you want to do. I spoke to a developer who wanted to create a new Tor connection to an electrum server for each address so the addresses couldn't be easily linked and to run this slowly in the background. It would be really difficult to implement WalletSync with such a strategy. Another example where WalletSync isn't the right fit is the Sensei (opens new window) project which uses BDK but incrementally updates the database whenever new information comes in from the blockchain.

Even if syncing all addresses at the same time is roughly what you want to do WalletSync still gets in the way since it defines whether you do it synchronously or asynchrononusly. Applications can control this through bdk's async-interface feature flag which internally changes the trait definition through macros. Another annoyance is that when using async-interface the future that gets returned from WalletSync cannot be Send (opens new window) because of how Wallet handles database mutability internally, meaning you can't spawn the future into a new thread.

# A general syncing mechanism

So what is the most general syncing mechanism that solves these problems? These are the things I think it has to do regardless of where the blockchain data comes from or how it's stored:

  1. Generate and store addresses.
  2. Index transaction data, e.g. transaction outputs we own, when/if they were spent, etc.
  3. Keep track of which addresses have been given out and which have been used.
  4. Be able to "roll back" our view of the above data if a re-org makes some of it stale.
  5. Keeping track of transactions related our addresses in our mempool.

Let's talk about how to implement a mechanism that does all that.

# How to store and index transactions

Different persistent storage backends have different APIs and their own indexing strategies. That's why the Database (opens new window) trait exists in BDK, to make a clean API to the different storage engines. It's important to note that the database in BDK only holds public data that could always be retrieved from the chain. It's just a cache. Despite this we support different backends. Right now it is a lot of work to add a new index to the data since you have to add it to every backend and you might have to apply schema changes (we still don't have a standard approach to this (opens new window)).

Thomas Eizinger suggested (opens new window) doing everything in memory and only writing to persistent storage when it was convenient. It took me some time but I came around to this idea. It would allow us to get rid of the Database trait (at least at the bdk_core level) and greatly simplify what the persistent storage layer has to do. Whenever the data is loaded from persistent storage we can just do the indexing in memory and present it to the application.

But wait! Wouldn't this mean we'd use way more memory than we need to? Yes but memory is cheap. Consider that if we say the average transaction size is 300 bytes then with all our indexes each transaction might cost 1kb of memory (pessimistically). This means we could index one thousand transactions in a single megabyte! My iPhone has 4gb of memory so it could index a million transactions with plenty of memory to spare. But what if some users can't afford an iPhone? Then they also couldn't have afforded to have made a million Bitcoin transactions! But what about memory constrained devices like hardware wallets!? Those devices typically don't store and retrieve transactions. They're usually just signing devices. Perhaps one day someone will build a memory constrained device that needs to do this work but until then I think this is a fine approach to take.

For now I'm calling this thing that does the in-memory indexing of transactions related to a single descriptor a DescriptorTracker. Here's a diagram that communicates how I imagine it relates to the other components.

# Rolling back, rolling forward and syncing to disk

State changes in blockchains are clearly delineated. They all happen in blocks! Every view of the blockchain, whether you're getting it through compact block filters, an electrum server or something wacky like a utreexo bridge will have a concept of blocks and transactions in them. For a wallet we only need a very sparse view of the blockchain that includes at which block a set of transactions existed. That way, if a block disappears we know that all those transactions might disappear too.

With bdk_core I want to introduce the concept of a checkpoint, which is a block height and hash and a set of txids that were present at that height but not present in the previous checkpoint. In this way we create an append-only data structure that can easily be rolled back to a previous height if there is a re-org. After rolling back we can then roll forward and apply the new blocks.

Here's an example of how this idea works:

There are a few edge cases I'd like to cover:

  1. What if when gathering new data from the chain to update a DescriptorTracker we find an old transaction that belongs to an earlier checkpoint that we had missed form our earlier syncs?
  2. What if when we go to write to persistent storage from a DescriptorTracker we find that it has some transactions the tracker doesn't? Should we try and reconcile the two sets of transactions?

I think the correct approach is to treat the chain data as the source of truth for the DescriptorTracker and the DescriptorTracker as the source of truth for persistent storage. That is in the case of (1) we should just rollback the DescriptorTracker and insert the old but recently discovered transaction in the right place. In the case of (2) we should roll back the persistent storage to the point where it differs and apply changes from there. This implies that you should only keep one instance of a DescriptorTracker for a descriptor in your application and only update persistent storage by first applying the changes to the tracker.

# Examples

Here are some examples of what I think this may end up looking like in code. Keep in mind that if this looks complicated it will probably be more complicated in practice! This doesn't mean that we can't create simplifying abstractions and tools around these primitives to cover common policies. I hope we can implement Wallet with DescriptorTrackers internally.

# Doing an initial sync of a descriptor that may already contain coins

When we first sync a descriptor that may already contain coins we want to iterate over all the scripts of the wallet and then stop if there's a big enough gap (e.g. 20). In this example we use an stateless esplora-like API (opens new window).

// create a descriptor tracker the external addresses of a BIP86 key
let mut tracker = DescriptorTracker::new("tr([73c5da0a/86'/0'/0']xpub6BgBgsespWvERF3LHQu6CnqdvfEvtMcQjYrcRzx53QJjSxarj2afYWcLteoGVky7D3UKDP9QyrLprQ3VCECoY49yfdDEHGCtMMj92pReUsQ/0/*)");

let esplora = bdk_esplora::Client::new();
let update = esplora.fetch_related_transactions(bdk_esplora::Params {
   // iterate over all addresses in a descriptor
   scripts: Some(tracker.iter_scripts()),
   // stop if you find a gap of 20 unused addresses
   stop_gap: Some(20),
   ..Default::default()
}).await?;

tracker.apply_update(update)?;

// now we want to persist this disk
let db_update = tracker.generate_update(Params {
    start_checkpoint: None,
});

// Note that the db_update type is the same as the `update` above.
my_db.apply_update(db_update);

# Doing a sync of a wallet after you already have sync'd

Now imagine you just want to check if any UTXOs in your wallet have been spent. In this case we've already sync'd before so we need to load that data into the tracker from disk first (rather than going straight to the blockchain). Then we just ask esplora for transactions related to these transaction outputs.

// create a descriptor tracker the external addresses of a BIP86 key
let mut tracker = DescriptorTracker::new("tr([73c5da0a/86'/0'/0']xpub6BgBgsespWvERF3LHQu6CnqdvfEvtMcQjYrcRzx53QJjSxarj2afYWcLteoGVky7D3UKDP9QyrLprQ3VCECoY49yfdDEHGCtMMj92pReUsQ/0/*)");

let init_update = my_db.generate_update(Params {
    checkpoint: None
});

// get up to speed with what was on disk.
tracker.apply_update(init_update);
// get the latest checkpoint
let checkpoint = tracker.get_checkpoint(0);

let esplora = bdk_esplora::Client::new();

// Fetch transactions spending any utxos we have
let update = esplora.fetch_related_transactions( bdk_esplora::Params {
   checkpoint: Some(checkpoint),
   tx_outs: Some(tracker.iter_unspent()),
   ..Default::default()
}).await?;

match update {
   Ok(update) => {
       tracker.apply_update(update)?;
       // now we want to persist this disk
       let db_update = tracker.generate_update(Params {
           // this call could fail if tracker no longer has this checkpoint.
           // In this case we'd ask persistent_storage for an earlier checkpoint and try again.
           start_checkpoint: persistent_storage.get_checkpoint(0),
       });

       persistent_storage.apply_update(db_update);
   }
   Err(bdk_esplora::Error::StaleCheckpoint) => {
      // here we should call fetch related transactions with an earlier checkpoint.
      // In practice this logic will be called in a loop
   }
}

# Updating state when you get the data in real time

If you have an event based view of the blockchain that feeds you block connected or block disconnected events then I imagine the API would look something like this. There's quite a bit left out here but I hope you get the idea.

// create a descriptor tracker the external addresses of a BIP86 key
let mut tracker = DescriptorTracker::new("tr([73c5da0a/86'/0'/0']xpub6BgBgsespWvERF3LHQu6CnqdvfEvtMcQjYrcRzx53QJjSxarj2afYWcLteoGVky7D3UKDP9QyrLprQ3VCECoY49yfdDEHGCtMMj92pReUsQ/0/*)");


let blockchain_events = { /* get a Stream of blockchain block connected/disconnected events */ };

loop {
    let blockchain_event =  blockchain_events.next().await;
    match blockchain_event {
        BlockChainEvent::Connected(new_block) => {
            match tracker.apply_block(new_block) => {
                Ok(modified) => if modified {
                    // update persistent storage from tracker
                }
                Err(ApplyBlockError::OutOfOrder) => {
                    // the block event we got was not the next block we expected.
                    // How to recover from this will depend on the application and block source
                }
           },
           BlockchainEvent::Disconnected((disconnected_height, disconnected_hash)) => {
              // this might invalidate a checkpoint
              tracker.disconnect_block(disconnected_height, disconnected_hash);
              // Now apply to persistent storage
           }
        }
    }
}

# Feedback

The best way to give feedback on this would be to comment on the pull request (opens new window) for this blog post. Thanks in advance.