Skip to content

Shard

The Shards together store and update the state of the replicated state machine and together are a component of the Execution Engines. They provide Executors with input data and update the state according to the results of Executors' computations.

Different shards may be on different physical machines.

Each shard is responsible for a set of KVSKeys and these sets are disjoint for different shards. For each of the keys that a shard is responsible for, the shard maintains a (partially-ordered) timeline of Timestamps of transaction candidates that may read or write to keys. Shards also keep a history of data written by each TransactionCandidate to each key. This is multi-version concurrent storage.

State (of the shard)

For each Worker Engine, the Shard maintains:

  • A Timestamp, such that all write lock requests1 for transaction candidates with earlier timestamps that this worker curates have already been received. Together, these timestamps represent heardAllWrites.
  • Another Timestamp, before which the Shard will receive no further read requests from this Worker Engine. For WorkerEngine, this cannot be after the corresponding write Timestamps. We will also maintain these from each Read Backend worker. Together, these represent heardAllReads.

For each key (assigned to this Shard):

  • A set of time‍stamps of known transaction candidates that read and/or write that key, and for each, some subset of:
  • A value written to that key at that time‍stamps by that TransactionCandidate using a KVSWrite message
  • A marker indicating that this TransactionCandidate may (or will) write to this key, but this Shard has not yet received a corresponding KVSWrite message.
  • A marker indicating that this TransactionCandidate will read this value, and an ExternalIdentity corresponding to the relevant Executor. This marker is only stored so long as the Shard doesn't know the value. When this value is determined, this Shard must remove this marker and send a KVSRead message to the Executor.
  • A marker indicating that this TransactionCandidate may read this value, and an ExternalIdentity corresponding to the relevant Executor. If the Executor sends a KVSReadRequest for this key, the Shard updates this marker to a "will read" marker.
  • If a Timestamp has no corresponding markers or values written, we don't have to store it.
  • If a value written is before heardAllReads, and there are no pending reads or writes before it, then we can remove all earlier values written.

Additionally, the Shard maintains:

  • A complete copy of the DAG structure produced by the Mempool Engines. This includes a set of all NarwhalBlockHeaders. For Timestamps before SeenAllRead, if there are no keys with a pending read or write before that Timestamp, we can delete old DAG structure.
  • A complete copy of the sequence of Anchors chosen by Consensus Engine. This is a sequence of consensus decisions. For Timestamps before heardAllReads, if there are no keys with a pending read or write before that Timestamp, we can delete old anchors.

Shard Optimizations

We want to execute each TransactionCandidate (evaluate the executor function in order to compute the data written) using the idea of serializability: each TransactionCandidate's reads and writes should be as if they were executed in the total order determined by the mempool (and [[Consensus

However, we want to compute concurrently as possible, for minimum latency. We do this using a set of optimizations.

Optimization: Per-Key Ordering

Per-key ordering (see web version for animation)

Mempool and consensus provides ordering information for the time‍stamps. Thus, relative to each key, transaction candidates can be totally ordered by the Happens Before relationship. With a total ordering of transaction candidates, Shards can send read information (KVSReads) to Executors as soon as the previous TransactionCandidate is complete. However, transaction candidates that access on disjoint sets of keys can be run in parallel. In the diagram above, for example, transaction candidates c and d can run concurrently, as can transaction candidates e and f, and transaction candidates h and j.

Optimization: Order With Respect To Writes

Order with respect to writes (see web version for animation)

In fact, Shards can send read information to an Executor as soon as the previous write's TransactionCandidate has completed (sent a KVSWrite). All Shards really need to keep track of is a total order of writes, and how each read is ordered with respect to writes (which write it precedes and which write preceded it). As soon as the preceding write is complete (the Shard has received a KVSWrite), the reads that depend on it can run concurrently. There are no "read/read" conflicts. In the diagram above, for example, transaction candidates a and b can run concurrently.

Optimization: Only Wait to Read

Only wait to read (see web version for animation)

Because we store each version written (multi-version concurrent storage), we do not have to execute writes in order. A Shard does not have to wait to write a later data version to a key just because previous reads have not finished executing yet. In the diagram above, for example, only green happens-before arrows require waiting. transaction candidates a, b, c, and j can all be executed concurrently, as can transaction candidates d, e, and i.

Optimization: Execute With Partial Order

Some mempools, including Narwhal, can provide partial order information on transactions even before consensus has determined a total order. This allows the Ordering Machine to execute some transactions before a total ordering is known. In general, for a given key, a shard can send read information to an executor when it knows precisely which write happens most recently before the read, and that write has executed.

heardAllWrites

In order to know which write happens most recently before a given read, the Shard must know that no further writes will be added to the timeline before the read. Mempool and consensus should communicate a lower bound on timestamps to the Shards, called heardAllWrites. The Shard is guaranteed to never receive another KVSAcquireLock with a write operation and Timestamp before heardAllWrites. In general, a Shard cannot send a KVSRead for a Timestamp unless the Timestamp is before heardAllWrites. heardAllWrites consists of a TxFingerprint from each worker engine such that the worker engine is certain (based on KVSLockAcquireds) that the Shard has already seen all the KVSAcquireLocks it will ever send at or before that TxFingerprint.

This can be on a per-key basis or simply a global lower bound. Occasionally, heardAllWrites should be updated with later timestamps. Each round of consensus should produce a lower bound for heardAllWrites, but the mempool may already have sent better bounds. Each Shard must keep track of heardAllWrites on each key's multi-version timeline.

Transactions (like transaction j in the diagram below) containing only write operations can execute with a timestamp after heardAllWrites, but this simply means calculating the data they will write. Since that does not depend on state, this can of course be done at any time.

heardAllReads

We want to allow Typhon to eventually garbage-collect old state. mempool and consensus should communicate a lower bound timestamp to the execution engine, called heardAllReads, before which there will be no more read transactions send to the execution engine. Occasionally, heardAllReads should be updated with later timestamps. Each Shard must keep track of heardAllReads on each key's multi-version timeline, so it can garbage-collect old values.

Execute with partial order (see web version for animation)

In the example above, our happens-before arrows have been replaced with may-happen-before arrows, representing partial ordering information from the mempool. Note that not all transactions can be executed with this partial order information.

Conflicts

There are three types of conflicts that can prevent a transaction from being executable without more ordering information.

  • Write/Write Conflicts occur when a shard cannot identify the most recent write before a given read. In the diagram above, transaction e cannot execute because it is not clear whether transaction b or transaction c wrote most recently to the yellow key.

  • Read/Write Conflicts occur when shard cannot identify whether a read operation occurs before or after a write, so it is not clear if it should read the value from that write or from a previous write. In the diagram above, transaction g cannot execute because it is not clear whether it would read the data written to the blue key by transaction d or transaction i.

  • Transitive Conflicts occur when a shard cannot get the data for a read because the relevant write is conflicted. In the diagram above, transaction h cannot execute because it cannot read the data written to the yellow key by transaction g, since transaction g is conflicted.

As the mempool and consensus provide the execution engine with more and more ordering information, and the partial order of timestamps is refined, all conflicts eventually resolve. In the diagram above, suppose consensus orders transaction g before transaction i. The Read/Write conflict is resolved: transaction g reads the data transaction d writes to the blue key. Then the transitive conflict is also resolved: transaction h will be able to execute. -->

Optimization: Client Reads as Read-Only Transactions

Client reads as read-only transactions (see web version for animation)

With the above optimizations, transactions containing only read operations do not affect other transactions (or scheduling) at all. Therefore, they can bypass mempool and consensus altogether. Clients can simply send read-only transactions directly to the execution engine (with a label and a timestamp), and if the timestamp is after heardAllReads, the execution engine can simply place the transaction in the timeline of the relevant shards and execute it when possible. In the diagram above, transaction f is read-only.

If client reads produce signed responses, then signed responses from a weak quorum of validators would form a light client proof.

Shard Incoming Messages

Shards receive and react to the following messages:

KVSAcquireLock

Purpose

Inform the shard about keys that a transaction may/will read and/or write, at a transaction fingerprint.

Structure

Field Type Description
lazy_read_keys KVSKey set Keys this transaction may read (only send values read in response to KVSReadRequests)
eager_read_keys KVSKey set Keys this transaction will read (send values read as soon as possible)
will_write_keys KVSKey set Keys this transaction will write. Future reads are dependent only on the KVSWrite for this TxFingerprint.
may_write_keys KVSKey set Keys this transaction may write. Future reads are dependent on the KVSWrite for this TxFingerprint, or, if that has a None, the previous value.
curator ExternalIdentity the Worker Engine in charge of the corresponding transactions
executor ExternalIdentity the Executor for this TransactionCandidate
timestamp TxFingerprint specifies the transaction affiliated with these locks.

The lazy_read_keys and eager_read_keys may not overlap. In the same way, will_write_keys and may_write_keys must be disjoint. There must be one KVSAcquireLock per Shard per TxFingerprint: for a given Shard and TxFingerprint, all information is to be provided in totality or not at all.1

Note that future versions may use some kind of structured Keys to encode "Sets" containing infinitely many Keys. For V1, however, simple HashSets or similar are fine.

Effects

  • The Shard stores the respective "locks" for all keys in its timeline.
  • these are the "markers" described in Shard State.
  • The eager_read_keys will be served as soon as possible (by sending KVSRead-messages to the executor).
  • The Shard immediately informs the curator that the locks are acquired, using a KVSLockAcquired message, so the curator can prepare UpdateSeenAll messages.

Triggers

KVSReadRequest

Purpose

Informs the Shard about a new read request, which happens in either of the following cases:

  • An Executor has determined that it actually needs the value at some key for which it has a lazy read (a may_read in the TransactionLabel of the TransactionCandidate). Now the executor is requesting that value from the Shard that stores it.
  • A Executor has finished and does not need the value for some key for which it has a lazy read (a may_read in the TransactionLabel).

Structure

Field Type Description
timestamp TxFingerprint we need the value at this logical timestamp
key KVSKey the value corresponds to this key
actual bool true iff we actually want a response

If actual is false, this just means that there is no dependency on this key in the current execution.

Effects

A Shard should delay processing a KVSReadRequest until it has completed processing KVSAcquireLock for the same timestamp.

Then, if actual is false, the Shard is done reading the value, and can remove the may read marker from state.

If actual is true, the Shard replaces the may read marker with a will read marker. If the Shard knows the unique previous value written before this timestamp, it sends that value in a KVSRead to the Executor and removes the will read marker from state. Otherwise, future KVSWrites and/or UpdateSeenAlls will identify this unique previous value written, and trigger the KVSRead.

Triggers

  • to Executor: KVSRead if the Shard has determined the unique value written prior to this "lock" then send a KVSRead-message to the relevant Executor to inform them of the value

KVSWrite

Purpose

Informs the Shard about a new write request, which happens in either of the following two cases:

Structure

Field Type Description
timestamp TxFingerprint the logical time at which we are writing this data.
key KVSKey the key used. With fancy hierarchical keys or suchlike, we could even specify a range of keys
datum KVSDatum option the new data to be associated with the key. No datum should only be used in a "may_write," and means don't change this value

Effects

A Shard should delay processing a KVSWrite until it has completed processing KVSAcquireLock for the same timestamp.

If the datum is None, then remove the may write marker from this timestamp in state. Any reads waiting to read what is written here must instead read from the previous write. - One way to accomplish this is to copy the previous write as a "value written" at this timestamp in state.

If datum is occupied, then remove the may write or will write marker from this timestamp in state, and record the value written at this timestamp in state.

This may trigger a KVSRead if there are any will read markers for which this timestamp is the unique previous write.

Triggers

UpdateSeenAll

  • from Mempool Engines

Purpose

In order to actually serve read requests, the Shard needs to know that it will not receive more write requests before a certain timestamp. These are in general broadcast to all Shards.

It is important that the Worker Engine has received KVSLockAcquired-messages for all KVSAcquireLocks it has sent (or will ever send) at or before the timestamp. In other words, shards know about all possible read and write requests of TransactionCandidates for which the worker is curator and may come earlier.

Todo

rephrase the above paragraph

Each worker engine only needs to send the Shard Engine UpdateSeenAll messages concerning worker-specific ordering (batch number and sequence number within the batch). This means that each Shard Engine needs to hear from every Worker Engine periodically to be sure it is not waiting for any transactions. From there, the Shard uses TimestampOrderingInformation about the Narwhal DAG and Consensus to fill in a total order.

Structure

Field Type Description
timestamp TxFingerprint represents a the position in the total order (in V1)
write bool seen all read and seen all write can (and should) be separate.

For V1, we only care about write = true because we don't garbage collect and assume multi-version storage. From V2 onward, the Shard is keeping additional ordering information and we do have garbage collection protocols.

Effects

Shards can now identify the unique previous write prior to each read at or before this timestamp.

If that unique previous write has a value written, and the read is marked will read, they can send a KVSRead with that value to the relevant Executor.

Triggers

  • to Executor: KVSRead for each will read for which we have established a unique previous write value send a KVSRead message to the relevant Executor

TimestampOrderingInformation

Purpose

While each transaction comes with a Timestamp, the shards do not actually know the order of those timestamps until the DAG is built, and consensus decisions are made. This message represents the mempool communicating (partial) timestamp ordering. These are broadcast to all shards.

Structure

Todo

One way to convey this is to include the entire DAG structure (albeit without the transaction contents of each worker batch). For now, I do not know what the internal structure of this message looks like.

Todo

check whether, ① worker-timestamp (= tx fingerprint), ② primary-timestamp (= pure DAG structure based on blocks/headers), ③ consensus-timestamp (= total order) are sufficiently many cases or we need yet another intermediate gradual step of ordering. E.g., does is make sense to also take into account local headers that (without integrity votes).

Effects

As shards learn more ordering information, they can finally complete reads (since they learn which writes most recently occurred).

Triggers

  • to Executor: KVSRead for each locked key for which we have established a unique write value, send a KVSRead message to the appropriate Executor

AnchorChosen

Purpose

Inform shards about the most recently decided value by the consensus.

Structure

Field Type Description
chain_id ChainId the chain Id
learner Learner learner (in V2 this is always \(\red\alpha\))
height Height height of the block
anchor NarwhalBlock the value that consensus decided upon

Effects

The shard learns more ordering information. In particular, with this and enough TimestampOrderingInformation messages, it should be able to order all transactions before the new anchor.

Once we have enough ordering information to establish the unique write preceding a key on which there is a read lock, and we have a value for that write, we can send that value to the relevant Executor.

Triggers

  • to Executor: KVSRead for each locked key for which we have established a unique write value, send a KVSRead message to the appropriate Executor.
(Wiki) links on this page

  1. For the purpose of this discussion, we call a write lock request a KVSAcquireLock message for a key for which a write request will or may be issued.