rewrite of sei-load generator and sender.

[pompon0]

Primary reason for the rewrite is support for non-happy paths for autobahn loadtesting - rejected transactions/not successfully executed/etc. The idea is that if sending a tx fails, we need to check the account nonce before sending the next transaction. Note that this affects only accounts from the long-lived pools - accounts generated for the sake of sending just 1 transaction (aka "new accounts") do not benefit from this extra logic.

The nonces, as well as the txs waiting to be sent are managed by TxsQueue, which ensures that:

• inserted (generated) transactions have expected nonces
• transactions per account are sent sequentially (otherwise they will be rejected by autobahn)

TxsQueue is managed by ShardedSender.

Additionally

• removed unnecessary abstractions in both sender and generator logic
• made Generator.Run accept the sender which exposes Nonce(), Flush() and Send() calls
• added Generator <-> Sender integration tests
• removed open/closed loop logic (might need to be reintroduced later if needed)
• removed utils/rng in favor of just using a single Rand synchronously (we don't have other use cases). If concurrent Rand usage is needed, a SplitRng method was introduced (slower but robust alternative to SplitMix).

[cursor]

PR Summary

High Risk
Core load-generation and submission path is redesigned (nonce recovery, queue semantics, removal of open-loop scheduling), which can change throughput, replay behavior, and failure handling in production load tests.

Overview
Rewrites the load-test generator and sender so failed sends on long-lived accounts can recover via on-chain nonce checks, and strips several layers of indirection.

Sender: The dispatcher, open/closed-loop scheduler, and sharded worker queues are gone. ShardedSender feeds a new TxsQueue that enforces per-address nonce order, caps backlog with MaxInFlight, and on RPC failure for tracked senders refetches nonce and **Reset**s queued txs. One ethClient dials all endpoints and picks a client by address hash; sends run in spawned goroutines after the shared rate limiter.

Generator: Generator.Run(ctx, rng, txSender) drives the campaign: weighted scenarios are expanded and shuffled once, then each tick pulls Nonce from the sender, builds txs with an explicit *rand.Rand, and **Send**s. Prewarm lives on Generator.Prewarm. Gas/distribution sampling no longer uses SetStream / utils/rng—callers pass Rand into SampleIndex / GenerateGas.

Other: Scenarios return signed Transaction + errors instead of LoadTx; FundAccounts takes a flat address list; settings Validate only requires MaxInFlight > 0 (arrival-model checks removed). Large test suites for the old scheduler/dispatcher are deleted; integration tests cover generator ↔ sender nonce rewind behavior.

^{Reviewed by Cursor Bugbot for commit a28e7c0. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Dropped txs look successful

Medium Severity

When TxsQueue.Push rejects a transaction for a nonce mismatch, it logs and returns nil. Generator.Run treats that as a successful send and immediately builds another transaction, which can spin at full speed without making progress until context cancel if the queue and generator stay desynchronized.

Additional Locations (1)

• generator/generator.go#L195-L198

 

^{Reviewed by Cursor Bugbot for commit 91db965. Configure here.}

[pompon0]

WAI

[cursor]

Buffer size no longer bounds queue

Medium Severity

NewShardedSender sizes TxsQueue only from Settings.MaxInFlight. Settings.BufferSize and LoadConfig.TotalQueueSize() are no longer used, so existing configs that tune bufferSize/--buffer-size do not limit queue depth or memory as before.

 

^{Reviewed by Cursor Bugbot for commit 34ba3fe. Configure here.}

[cursor]

Worker queue metric disabled

Low Severity

The worker_queue_length observable gauge callback no longer records observations: the ShardStats loop is commented out and only assigns _ = ss. Scrapers always see an empty series after the sender rewrite removed per-shard queues.

 

^{Reviewed by Cursor Bugbot for commit 34ba3fe. Configure here.}

[cursor]

Stale nonce before async send

Medium Severity

Run reads txSender.Nonce, builds and signs the tx, then calls Send without holding sender/queue state. A concurrent send failure and Reset can advance the queue nonce before Push, producing signed txs the queue rejects.

 

^{Reviewed by Cursor Bugbot for commit b83e882. Configure here.}

[pompon0]

Rejection is silent and it is expected, since we send the txs asynchronously.

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

There are 5 total unresolved issues (including 4 from previous reviews).

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit a28e7c0. Configure here.}

[cursor]

Workers setting no longer applied

Medium Severity

The send path now runs a single ShardedSender background task and no longer uses Settings.TasksPerEndpoint (workers) or LoadConfig.GetNumShards, even though those settings remain in config, CLI, and tests. Raising --workers does not increase RPC parallelism as documented.

 

^{Reviewed by Cursor Bugbot for commit a28e7c0. Configure here.}