Transaction Pipeline

The transaction pipeline handles the full lifecycle from creation to finalization.

Overview

User/SDK
  │ build + sign (Ed25519)
  ▼
JSON-RPC (savitri_sendRawTransaction)
  │ hex decode → bytes_to_raw_tx
  ▼
Mempool Admission
  │ stateless guards → stateful guards
  ▼
SIMD Batch Scoring
  │ fee * 0.7 + class_priority * 0.3
  ▼
Block Proposer (drain_for_block_production)
  │ top-N by score, nonce-ordered per sender
  ▼
Block Production
  │ compute tx_root, state_root
  ▼
Gossipsub Broadcast (/savitri/block/1)
  │
  ▼
BFT Voting (2f+1 quorum)
  │ Block Acceptance Certificate
  ▼
RocksDB Commit
  │ CF_BLOCKS, CF_TRANSACTIONS, CF_ACCOUNTS
  ▼
Finalized

Transaction Format

Canonical Signing Message

from (hex bytes) || to (hex bytes) || amount (u64 LE) || nonce (u64 LE) || fee (u128 LE)

This message is SHA-256 hashed, then signed with Ed25519.

Wire Format (raw_tx_hex)

from_hex_bytes || to_hex_bytes || amount_le_u64 || nonce_le_u64 || fee_le_u128 || pubkey (32B) || signature (64B) || data (optional)

Submitted as hex-encoded string to savitri_sendRawTransaction.

Mempool Admission

Stateless Guards (no storage access)

1. Signature format validation (64-byte ed25519)
2. Public key format validation (32-byte ed25519)
3. Signature verification against message hash
4. Transaction size limits

Stateful Guards (storage access required)

1. Nonce check: tx.nonce >= account.nonce (gap tolerance up to 5000)
2. Balance check: account.balance >= tx.amount + tx.fee
3. Quota check: Per-class and per-sender limits

Transaction Classes

Class	Priority	Max Quota
FederatedUpdate	1.0 (highest)	50,000
Financial	0.8	50,000
IoTData	0.5	100,000
Governance	0.7	50,000

Quota System

• Per-class limits: Financial 50K, IoT 100K, Global 100K
• Per-sender cap: 512 pending transactions per sender
• Fair drain: drain_fair_batch() respects quotas and calls record_removal()

SIMD Scoring

The dispatcher scores transactions using SIMD-optimized batch computation:

score = fee_normalized * fee_weight + class_priority * class_weight

Default weights: fee_weight = 0.7, class_weight = 0.3.

SIMD Implementation

Architecture	Instruction Set	Lanes	Threshold
x86_64	AVX2 + FMA	4 doubles	32 TXs
ARM (aarch64)	NEON	2 doubles	32 TXs
Fallback	Scalar	1	Always

For batches < 32 transactions, scalar computation is used (SIMD setup overhead exceeds benefit).

Score Cache

LRU cache avoids recomputing scores for transactions that remain in the mempool across blocks.

Adaptive Weights

With the adaptive_weights feature flag, weights adjust dynamically based on:

• Fee distribution in the mempool
• Class diversity
• Historical throughput

Block Production

The elected proposer:

1. Calls drain_for_block_production(max_block_txs)
2. Gets top-N transactions ordered by score
3. Runs final_validation with pending_nonces tracking (prevents stale nonce rejection)
4. Computes transaction root: "TXv1" leaf tag, rolling accumulator
5. Computes state root: "STATEv1-LE" seed, lexicographic snapshot
6. Signs block with Ed25519
7. Broadcasts via gossipsub

Nonce Handling

• final_validation tracks pending_nonces HashMap for drained-but-uncommitted TXs
• Cold-start fix: Accepts lowest available nonce when storage_nonce=0 AND pending_nonces=0
• Nonce reset: TX generator resets when local nonce > storage_nonce + 200

Storage

Committed blocks are stored in RocksDB column families:

Column Family	Key	Value
CF_BLOCKS	height (u64 LE)	Block (bincode)
CF_TRANSACTIONS	tx_hash	Transaction (bincode)
CF_ACCOUNTS	address (32 bytes)	Account (bincode)
CF_METADATA	"chain_head"	Latest block (bincode)

Empty accounts (balance=0, nonce=0) are never persisted.