SIMD-0083: Relax Entry Constraints #83

apfitzge · 2023-11-02T06:05:55Z

The second in a series of SIMDs aimed at the relaxation of constraints blocking us from (potentially) having asynchronous execution.
This proposal does not lock Solana into moving towards asynchronous execution.
This proposal stands on its own to simplify the protocol and give more flexibility in implementation of block-production.

lheeger-jump · 2023-11-02T17:55:20Z

proposals/0083-relax-entry-constraints.md

+## Security Considerations
+
+None
+
+## Drawbacks
+
+None


This change breaks consensus in a pretty fundamental way, which is both a security issue and has some drawbacks.

second that

@lheeger-jump can you please elaborate how will it break the consensus.

@Huzaifa696 - Per the below snippet, a self-conflicting entry is currently considered invalid, and will raise an error:
https://github.com/solana-labs/solana/blob/230779d4590ac625e6710117987ef505c0cf71c6/ledger/src/blockstore_processor.rs#L606-L619

That error results in the slot getting marked dead by ReplayStage:
https://github.com/solana-labs/solana/blob/230779d4590ac625e6710117987ef505c0cf71c6/core/src/replay_stage.rs#L2782-L2786

If the "rules" are changed to allow a self-conflicting entry, then a slot with a self-conflicting entry will no longer error / result in dead slot.

Suppose that the change to allow self-conflicting entries first lands in v1.18. Nodes running v1.17 will mark slots with a self-conflicting entry dead whereas nodes running v1.18 will allow it. This is obviously problematic during an upgrade period where the stake is evenly distributed between v1.17 and v1.18. As such, the change to allow self-conflicting entries would have to sit behind a feature gate, and only get enabled once the entire cluster has adopted v1.18.

yes, I agree with the problem and the solution you explained.
In conclusion, the required changes for this SIMD to allow self-conflicting entries have to sit behind a feature gate.

So, as far as I understand, the flow is: approved SIMD -> feature gate issue -> PR (that implements the SIMD with a feature gate). Is that right? @steviez

So, as far as I understand, the flow is: approved SIMD -> feature gate issue -> PR (that implements the SIMD with a feature gate). Is that right?

I believe so, yes

So, as far as I understand, the flow is: approved SIMD -> feature gate issue -> PR (that implements the SIMD with a feature gate). Is that right? @steviez

@jacobcreech would know best.

@lheeger-jump can you please elaborate how will it break the consensus.

As @steviez mentioned, this changes the validation criteria for blocks, which affects whether they are rejected. It needs a feature gate which implies coordination with other clients.

approved SIMD -> feature gate issue -> PR

That would be ideal to avoid any work that may be thrown away due to never getting consensus on a SIMD.

lheeger-jump · 2023-11-02T17:56:31Z

proposals/0083-relax-entry-constraints.md

+## Detailed Design
+
+Currently, if a transaction entry within a block contains transactions that
+conflict, the entire block is marked invalid.
+The proposal is that this constraint is removed entirely, and entries are
+allowed to contain transactions that conflict with each other.
+


How are clients supposed to handle write contention? Is it sequential in the entry? There is a lot to consider here.

The execution in an entry is always sequential whether its execution or validation. So, we have to change the lock function in a way that allows conflicting txs in the same entry as they will be executed in a sequential manner. You can refer to this comment

It wouldnt hurt to add that the execution now becomes sequential in the transactions intra microblock.

This proposal is not making execution sequential within an entry. It has already been sequential, it is only that individual entries run in parallel (provided that they are not conflicting with each other).

@apfitzge can you please confirm this behavior.

Not currently sequential, and actually has no implied or guaranteed execution order during replay.
I will explicitly state that an entry has a sequential ordering, so for conflicting txs they must now be executed sequentially.

Added clarifying comments in f76e7d3.
The proposed behavior is to execute conflciting transactions sequentially, in the order they appear in the entry.

lheeger-jump · 2023-11-02T17:57:32Z

proposals/0083-relax-entry-constraints.md

+## Impact
+
+- The protocol is simplified
+- Block-production is more flexible
+


It has a large affect on consensus

While I don't disagree with your comments that a consensus altering change could fit many of the sections on this document (such as Impact / Security / Drawbacks), it could get repetitive to mention it multiple times. I'm wondering if mentioning it once is sufficient; maybe here given the prompt:

solana-improvement-documents/XXXX-template.md

Lines 64 to 67 in ffb3cca

## Backwards Compatibility *(Optional)*

Does the feature introduce any breaking changes? All incompatibilities and

consequences should be listed.

I suppose I was under the assumption that all SIMDs are consensus breaking. Is that not the case?

I was under the asssumption that a consensus change and feature-gating was implicit all proposals here; I can be explicit about that.

I suppose I was under the assumption that all SIMDs are consensus breaking. Is that not the case?

I was under the asssumption that a consensus change and feature-gating was implicit all proposals here; I can be explicit about that

I haven't been involved with enough SIMD's to know 😬

All feature gates should be consensus changing, SIMDs do not need to be. A SIMD can propose application standards or networking changes.

f76e7d3: Added comments about consensus breakage under security considerations.

steviez · 2023-11-03T03:29:13Z

proposals/0083-relax-entry-constraints.md

+This proposal is backwards compatible with the current protocol, since it only
+relaxes constraints, and does not add any new constraints.
+All previously valid blocks would still be valid.


Not to beat a dead horse, but Liam is correct about this being a consensus altering change.

All previously valid blocks would still be valid.

This is true. However, a block that was previously invalid might now be valid. Or stated another way, a block might be valid under this new set of rules but invalid under the old set. As with other consensus altering changes, this could create a problem during a network upgrade where say 60% is running with old rules / 40% with new rules. As such, this change would need to be rolled out with a feature gate.

See other comment here: #83 (comment)

However, a block that was previously invalid might now be valid.

Your statement is true, but this seems like a *forward compatibility issue to me.
After the activation of the feature, we don't have previously invalid blocks because they cannot, by definition, be in the history. Therefore all previous (backward) blocks are still valid with the changes

*forward compatibility issue to me.

Yep, agreed. I guess this gets back to whether a SIMD is implicitly thought of as a consensus breaking change or not. In any case, you're seemingly well aware of the implications of the change so marking this item as resolved, and we can figure out how explicit we need to be in stating consensus implications

Added a small comment that newly produced blocks may not be valid under the old protocol: f76e7d3

steviez · 2023-11-03T03:42:19Z

proposals/0083-relax-entry-constraints.md

+## Impact
+
+- The protocol is simplified
+- Block-production is more flexible
+


While I don't disagree with your comments that a consensus altering change could fit many of the sections on this document (such as Impact / Security / Drawbacks), it could get repetitive to mention it multiple times. I'm wondering if mentioning it once is sufficient; maybe here given the prompt:

solana-improvement-documents/XXXX-template.md

Lines 64 to 67 in ffb3cca

## Backwards Compatibility *(Optional)*

Does the feature introduce any breaking changes? All incompatibilities and

consequences should be listed.

godmodegalactus · 2023-11-03T11:42:46Z

proposals/0083-relax-entry-constraints.md

+The current constraint that transaction entries must not conflict with each
+other is unnecessary.
+It is a constraint, that made the current labs-client implementation easier,
+but is not required for the protocol to function correctly.


Should describe more what is intended behavior incase of transactions writing the same accounts.

@godmodegalactus currently, what happens is that before actual execution, all accounts of an entry are loaded and then the same state is used for execution. This behavior does not allow "writing the same accounts" as is.
But if we carry the temporary changed state as the execution proceeds sequentially in an entry and then commit the final state then there is no issue in writing the same accounts in an entry.

In an entry all the transactions should be able to run in parallel. If in an entry two transactions are loading same accounts, these transactions cannot be executed in parallel right. Even they will have to maintain ordering of the executing because if one is executed before another in replay then consensus will break. So i guess just explicitly mentioning these cases is important.

In the latest implementation, txs "within an entry" are not executed in parallel in block production or in replay, rather multiple entries run in parallel if they don't conflict with each other.

so, we are safe from the race condition scenario you mentioned

@godmodegalactus added comments in f76e7d3. Conflicting transactions within an entry must be executed sequentially, in the order they appear in the entry.

@Huzaifa696 current master still does rebatching of transactions though. Prior versions, which are the majority of the network, even do randomization on the order of txs within the non-conflicting entries.

talalim · 2023-11-09T05:35:30Z

proposals/0083-relax-entry-constraints.md

+conflict, the entire block is marked invalid.
+The proposal is that this constraint is removed entirely, and entries are
+allowed to contain transactions that conflict with each other.
+


We implemented and tested the required changes to enable self conflicting entries in our lab setup and this improves the performance a lot depending on how you schedule the transactions on every execution thread.
The changes include:

changing account locking function

carrying temporary account state within the execution with every transaction in an entry

As the functions changed are used by both the executor and validator, so it doesn't break any consensus and rules of the chain, and is backward compatible.
These changes can be implemented through a feature gate.

@apfitzge we can add these points to the detailed design if you think it's appropriate.

As the functions changed are used by both the executor and validator, so it doesn't break any consensus and rules of the chain, and is backward compatible.
These changes can be implemented through a feature gate.

This changes the structure of blocks and that directly affects consensus. In Firedancer today you would break validation criteria for a block and there would be a hard fork from Solana Labs.

@lheeger-jump is correct, this is a consensus breaking change. If you produce a block that contains entries with conflicting transactions, existing validators will mark it as invalid. For this reason, it needs to sit behind a feature, and only be activated when both validator clients have implemented the feature-gate and those versions have been rolled out to cluster(s).

@apfitzge we can add these points to the detailed design if you think it's appropriate.

This doesn't seem to be appropriate for a SIMD. SIMDs should be independent of any specific client's implementation. The issue in the labs repo you created would be a more appropriate place to share implementation-specific code and any results.

I've shared some implementation specifics on the issue—please take a look.

apfitzge · 2024-01-09T18:28:42Z

@lheeger-jump finally found time to update this. Added a list of the remaining constraints on individual entries.
Wasn't too sure how to list the bincode deserialization constraint, seems Richard took a shot at documenting that in the specs document: solana-foundation/specs#17

I think I've now hit all your requested changes, lmk if this is not the case.

0xSol · 2024-04-08T17:28:53Z

@lheeger-jump please review requested changes.

lheeger-jump

This looks great. Simple, concise and precise such that I can implement. Good work.

0xSol · 2024-04-09T18:16:12Z

@tao-stones could you or someone from Anza please review?

tao-stones

Am happy with this change

0xSol · 2024-04-09T21:15:29Z

@t-nelson Could you please review then comment or approve from Anza side?

lheeger-jump · 2024-04-14T18:46:56Z

Since we have an approval from Anza in the form of @tao-stones (please become an official reviewer via process in SIMD-0007), and no substantive comments since January, I am going to merge this PR so it does not go stale.

Thanks @apfitzge for your patience.

apfitzge added 2 commits November 2, 2023 12:36

Add XXXX-relax-entry-constraints.md

101e81b

Fill in SIMD-0083

fdcd8bd

apfitzge changed the title ~~SIMD-XXXX: Relax Entry Constraints~~ SIMD-0083: Relax Entry Constraints Nov 2, 2023

apfitzge marked this pull request as ready for review November 2, 2023 06:10

apfitzge mentioned this pull request Nov 2, 2023

Allowing entries containing transactions with conflicting accounts in Banking stage solana-labs/solana#33899

Closed

lheeger-jump requested changes Nov 2, 2023

View reviewed changes

steviez reviewed Nov 3, 2023

View reviewed changes

godmodegalactus reviewed Nov 3, 2023

View reviewed changes

talalim reviewed Nov 9, 2023

View reviewed changes

Explicitly specify consensus-break and sequential ordering

f76e7d3

apfitzge mentioned this pull request Nov 14, 2023

SIMD-0082: Relax Transaction Constraints #82

Closed

Huzaifa696 mentioned this pull request Nov 26, 2023

Allow self conflicting batches solana-labs/solana#34104

Closed

Terminology, explicit constraints

a009d13

apfitzge requested a review from lheeger-jump January 30, 2024 20:27

Clarify Tick entry terminology, and num hashes constraint

b153305

apfitzge mentioned this pull request Mar 13, 2024

SVM - requirement of load_and_execute_sanitized_transaction anza-xyz/agave#228

Open

lheeger-jump approved these changes Apr 9, 2024

View reviewed changes

tao-stones approved these changes Apr 9, 2024

View reviewed changes

lheeger-jump merged commit 44feb6a into solana-foundation:main Apr 14, 2024
2 checks passed

github-actions bot mentioned this pull request Apr 15, 2024

SIMD Updates - Mon Apr 15 02:28:40 UTC 2024 smartcontractkit/chainlink-solana#661

Closed

apfitzge deleted the relax-entry-constraints branch April 15, 2024 16:38

Huzaifa696 mentioned this pull request Apr 24, 2024

Allow Self-Conflicting Entries in execution anza-xyz/agave#1025

Open

This was referenced Apr 24, 2024

Feature Gate: SIMD-0083: Relax Entry Constraints anza-xyz/agave#1029

Open

SIMD-0083: update status and feature #147

Merged

Huzaifa696 mentioned this pull request Jul 24, 2024

SIMD#83 - Allow conflicting entries on replay side anza-xyz/agave#2256

Open

2501babe mentioned this pull request Oct 16, 2024

svm: allow conflicting transactions in entries anza-xyz/agave#3146

Closed

2501babe mentioned this pull request Nov 1, 2024

svm: interleave transaction validation and processing anza-xyz/agave#3404

Merged

2501babe mentioned this pull request Nov 14, 2024

svm: allow conflicting transactions in entries anza-xyz/agave#3453

Merged

	## Backwards Compatibility (Optional)

	Does the feature introduce any breaking changes? All incompatibilities and
	consequences should be listed.

SIMD-0083: Relax Entry Constraints #83

SIMD-0083: Relax Entry Constraints #83

Conversation

apfitzge commented Nov 2, 2023

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Huzaifa696 Nov 13, 2023 • edited Loading

Choose a reason for hiding this comment

apfitzge Nov 13, 2023 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

apfitzge Nov 3, 2023 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Huzaifa696 Nov 9, 2023 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Huzaifa696 Nov 9, 2023 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

apfitzge commented Jan 9, 2024

0xSol commented Apr 8, 2024

lheeger-jump left a comment

Choose a reason for hiding this comment

0xSol commented Apr 9, 2024

tao-stones left a comment

Choose a reason for hiding this comment

0xSol commented Apr 9, 2024

lheeger-jump commented Apr 14, 2024 • edited Loading

Huzaifa696 Nov 13, 2023 •

edited

Loading

apfitzge Nov 13, 2023 •

edited

Loading

apfitzge Nov 3, 2023 •

edited

Loading

Huzaifa696 Nov 9, 2023 •

edited

Loading

Huzaifa696 Nov 9, 2023 •

edited

Loading

lheeger-jump commented Apr 14, 2024 •

edited

Loading