Перейти к основному содержимому

Epoch Accounts Hash

Paraphrasing from https://github.com/solana-labs/solana/issues/26847

Background

Rent collection checks every account at least once per epoch. At each slot, a deterministic set of accounts (based on the pubkey range) is loaded, checked for rent collection, and stored back to the Accounts DB at this new slot.

Accounts are stored (rewritten) even if they are unchanged. This has a few positive effects.

  1. Once an account is rewritten, the previous version at an older slot is now dead. As entire slots and AppendVecs become full of only dead accounts, they can then be dropped/recycled.
  2. Each account rewritten due to rent collection is included in that slot's bank hash. Since the bank hash is part of what is voted on for consensus, this means every account is verified by the network at least once per epoch.

However, there is a big downside to rewriting unchanged accounts: performance. Storing accounts can be very expensive. And since accounts now are required to be rent-exempt, the majority of accounts rewritten due to rent collection are unchanged. What if unchanged accounts were no longer rewritten? This would minimally be a big performance win.

Problem

If rent collection no longer is rewriting unchanged accounts, we lose the two positive effects. Dealing with Positive Effect 1 (from above) will be handled by Ancient AppendVecs, and will not be discussed here. So how do we still get the security from Positive Effect 2? How can we still verify every account at least once per epoch, as part of consensus, but without rewriting accounts?

Proposed Solution

Perform a full accounts hash calculation once per epoch, and hash the result into a bank's hash. This will be known as the Epoch Accounts Hash, or EAH.

This retains the Positive Effect 2 from rent collection by checking every account at least once per epoch. Thus, any validators with missing, corrupt, or extra accounts will identify those issues within 1-2 epochs.

Implementation

Performing a full accounts hash takes a relatively long time to complete. For this reason, the EAH calculation must take place in the background.

In order for all the validators to calculate the same accounts hash, the calculation must be based on the same view of all accounts. This means the EAH must be based on a predetermined slot. This will be known as the start slot. The start slot is calculated as an offset into the epoch. This offset will be known as the start offset. Formally, the start slot is the first root greater-than-or-equal-to first slot in epoch + start offset.

Similarly, all the validators must save the EAH into a bank at a predetermined slot, and offset from the first slot of an epoch. This will be known as the stop slot and stop offset, respectively.

  • The start offset will be set at one-quarter into the epoch.
  • The stop offset will be set at three-quarters into the epoch.
  • For epochs with 432,000 slots, the start offset will be 108,000 and the stop offset will be 324,000.

These constants may be changed in the future, or may be determined at runtime. The main justifications for these values are:

  1. Do not start the EAH calculation at the beginning of an epoch, as the beginning of an epoch is already a time of contention and stress. There is no reason to make this worse.
  2. The bank to save the EAH into—the stop offset—should be sufficiently far in the future to guarantee all validators are able to complete the accounts hash calculation in time.
  3. The start offset should be after the rewarding interval (from Partitioned Inflationary Rewards Distribution). This ensures stake rewards have been distributed and stored into the accounts for this epoch.

Once the EAH calculation is complete, it must be saved somewhere. Since this occurs in the background, there is not an associated Bank that would make sense to save into. Instead, a new field will be added to AccountsDb that will store the EAH. Later, the bank at slot stop slot†¹ will read the EAH from AccountsDb and hash it into its own hash (aka bank hash).

EAH calculation will use the existing accounts background services (ABS) to perform the actual calculation. Requests for EAH calculation will be sent from bank_forks::set_root()†², with a new request type to distinguish an EAH request from a Snapshot request. Since the EAH will be part of consensus, it is not optional; EAH requests will have the highest priority in ABS, and will be processed first/instead of other requests.

†¹: More precisely, all banks where bank slot >= stop slot and parent slot < stop slot will include the EAH in their bank hash. This ensures EAH handles forking around stop slot, since only one of these banks will end up rooted.

†²: An EAH calculation will be requested when root bank slot >= start slot and root parent slot < start slot. This handles the scenario where validators call bank_forks::set_root() at different intervals.

Details

Snapshots

A snapshot contains all the state necessary to reconstruct the cluster as of a certain slot. A snapshot may then need to contain the EAH so that the stop slot can include it. Consider the following scenarios within an epoch where a snapshot is requested for slot X:

1. X >= first slot in epoch and X < start slot

Since the start slot has not been reached yet, there is nothing special to do in order to take a snapshot in this scenario.

2. X == start slot

The EAH must be included in the snapshot. Since the snapshot process always calculates the accounts hash, no additional calculations are required. The accounts hash calculation result will be used both to store in the snapshot as the EAH, and for the snapshot hash (which is used at load-time for verification).

3. X > start slot and X < stop slot

If a snapshot is requested to be created after the start slot but before the EAH calculation has completed, then it will be impossible to create a snapshot with the correct EAH. The snapshot process will wait until the EAH calculation has completed before proceeding.

4. X == stop slot

The EAH has been calculated for this epoch, and has been included in the stop slot bank. No further handling is required; the snapshot does not need to contain the EAH.

5. X > stop slot and X <= last slot in epoch

Same as (4).

Snapshot Verification

If a snapshot archive includes an EAH, we want to verify the EAH is correct at load time (instead of waiting until stop slot, which could be far in the future).

If the snapshot archive is for a slot within the calculation window†¹, then it must include an EAH. The snapshot hash itself will now also incorporate the EAH. In pseudo code:

if slot is in calculation window
let snapshot hash = hash(accounts hash, epoch accounts hash)
else
let snapshot hash = accounts hash
endif

Since loading from a snapshot archive already verifies the snapshot archive's hash against the deserialized bank, the EAH will be implicitly verified as well.

†¹: The calculation window is [start slot, stop slot), based on the epoch of the referenced Bank.

Corner Cases

Minimum Slots per Epoch

An EAH is requested by BankForks::set_root(), which happens while setting roots. The EAH is stored into Banks when they are frozen. Banks are frozen at least 32 slots before they are rooted. For the expected behavior, the EAH start slot really should be 32 slots before the stop slot. If the number of slots per epoch is small, this can result in surprising behavior.

Example: Assume there are 40 slots per epoch. The EAH start offset is 10, and the EAH stop offset is 30. When Bank 30 is frozen it will include the EAH in its hash. However, Bank 10 has not yet been rooted, so a new EAH has not been calculated for this epoch. This means Bank 30 will have included the EAH from the previous epoch in its hash.

Later, when Bank 10 is rooted, it will request a new EAH be calculated. If a snapshot is taken for Bank 12 (or any bank between 10 and 30), it will include the EAH from this epoch. If a node loads the snapshot from Bank 12, once it gets to freezing Bank 30, it will end up with a different bank hash since it included the EAH from this epoch (versus the other node's Bank 30 included the EAH from the previous epoch). Different bank hashes will result in consensus failures.

The above example is clearly bad. It can be observed that short epochs only occur (1) during warmup, or (2) in tests. Real clusters have much longer epochs (432,000 slots by default).

Tests can be fixed as needed; that leaves fixing warmup. Since warmup is transient, we disable EAH until slots-per-epoch is large enough. More precisely, we disable EAH until the calculation window is big enough. During warmup, slots-per-epoch doubles each epoch until reaching the desired number, so only a few epochs will skip EAH (which also is a small total number of slots).

Warping

Warping introduces corner cases into EAH because many slots may be skipped, including the entire range of start slot to stop slot.

Now, a new EAH will always be calculated based on the parent. Thus, if the warp slot is stop slot or greater (and the parent slot is less than stop slot), the warp bank will include this newly-calculated EAH. This is safe because warping cannot be used on a live cluster; only for a new cluster or tests/debugging. Therefore when the EAH was calculated is not germane.

For specific examples, refer to Appendix A.

Implementation Alternatives

Perform the EAH calculation in the foreground

The accounts hash calculation takes around 15 seconds (median, on Mainnet-Beta today). This is far beyond the slot time; this would be bad UX, and also decrease network stability.

Remove stop offset

Instead of having two offsets—one for start and one for stop—use a single offset for both. This delays when the EAH is saved into a Bank and voted on. The saved EAH is now the EAH from the previous epoch. This could work; would reduce the number of "special" slots from two to one. No significant advantages observed.

Send EAH requests when making a new bank, instead of a new root

When a bank is created, we don't yet know if it will be finalized until it is rooted, which could result in multiple EAH requests due to forking. This would be bad for performance.

Appendix A: All Warping Scenarios

To enumerate how EAH interacts with warping, refer to the following diagram for the scenarios below:

  +---------+-----------------+-----------+---------+-----------------+-----------+
| > < | > < |
| A > B < C | D > E < F |
| > < | > < |
+---------+-----------------+-----------+---------+-----------------+-----------+
| | | | | | |
v v v v v v v
epoch 1 start slot 1 stop slot 1 epoch 2 start slot 2 stop slot 2 epoch 3

parent slot: A, warp slot: A

No slots important to the EAH have been skipped, so no change in behavior.

parent slot: A, warp slot: B

An EAH calculation will be requested when the warp slot is rooted, and then will be included in the Bank at slot slot 1; behavior is unchanged.

parent slot: A, warp slot: C or D

The entire EAH range has been skipped; no new EAH calculation will have been requested for start slot 1. The warp slot will include the EAH calculated at the parent slot. This is different from the normal behavior.

parent slot: A, warp slot: E

Similar to A -> C, the warp slot will include the EAH calculated at the parent slot. This behavior appears different.

Then when the warp slot is rooted, a new EAH calculation will be requested, which will be included in stop slot 2. This behavior is expected.

parent slot: A, warp slot: F

Similar to A -> C, no new EAH calculation will be requested. The warp slot will include the EAH calculated at the parent slot. This is different from the normal behavior.

parent slot: B, warp slot: B

A new EAH will be calculated at parent slot, which will be included in stop slot 1, not the previously-calculated EAH from start slot 1. This behaior is different.

parent slot: B, warp slot: C or D

The warp slot will include the EAH calculated at the parent slot, not the previously-calculated EAH from start slot 1. This is different from normal behavior.

parent slot: B, warp slot: E

Similar to B -> C, the warp slot will include the EAH calculated at the parent slot, not the previously-calculated EAH from start slot 1. This is different from normal behavior.

And similar to A -> E, when the warp slot is rooted, a new EAH calculation will be requested, which will be included in stop slot 2. This behavior is expected.

parent slot: B, warp slot: F

Similar to B -> C, the warp slot will include the EAH calculated at the parent slot, not the previously-calculated EAH from start slot 1. This is different from normal behavior.