Deep in the heart of the imperial hinterland lies the Valley of Heavenly Craft, where nine world-renowned artisan guilds stand nestled along the mist-shrouded mountain terraces.

The Ironworks Ward forged hardened blades and massive clockwork gear trains; the Timber Ward shaped thousand-pace celestial chariots with intricate mortise-and-tenon framing; the Loom Ward spun damask silk linings as resilient as fine metal; and the Jade Ward polished frictionless mechanical bearings and rune-inscribed balance plates. The collective craftsmanship of the valley reached the pinnacle of mortal engineering. Whenever the imperial court or the grand trading leagues commissioned a monumental artifact, it invariably required the seamless cooperation of multiple guilds.

Consider, for example, the imperial commission for an armored “Celestial Steed Carriage.” The endeavor was gargantuan: it demanded three tempered leaf springs and oil-quenched steel axles from the Ironworks Ward, a fragrant sandalwood carriage frame assembled by the Timber Ward, and double-layered arrow-resistant raw silk upholstery stitched by the Loom Ward.

In earlier years, whenever the valley undertook such cross-guild undertakings, it relied entirely on impromptu, ad-hoc coordination among the master craftsmen.

Whenever an emissary arrived from the capital with a massive line of credit, the dispatch commissary would sprint to the Ironworks Ward, plunge an order stake into the foreman’s bench, and bellow: “Claim Forge Anvil No. 1 for Order A!” Next, the commissary dispatched a cohort of nimble-footed apprentices through miles of muddy mountain paths to the Timber Ward and Loom Ward, racing to claim the primary wood lathe and the master silk loom before anyone else could touch them.

This operational model governed the valley for decades, yet it eventually pushed the entire complex to the brink of collapse.

The primary disaster was cross-guild stalemate and gridlock.

One spring, the frontier command posted an urgent levy for eight hundred suits of iron lamellar armor, while the Ministry of Revenue simultaneously mandated one hundred river barges. The armor required the Ironworks Ward to forge breastplates before the Loom Ward could line them with hardened leather. The barges required the Loom Ward to weave heavy sails before the Ironworks could forge iron anchor chains.

The armor apprentices rushed in and claimed the great furnaces of the Ironworks, only to find the master loom in the Loom Ward already reserved for barge sails. Meanwhile, the barge overseers tenaciously held onto the looms, yet could not proceed because their apprentices were trapped waiting for an anvil in the Ironworks to free up. Neither faction would yield. Forge fires cooled, looms sat frozen in mid-stroke, and hundreds of master smiths and weavers stood glaring at one another with hammers and shuttles poised in vain. The standoff paralyzed the workshops for three days and three nights. Deadlines were blown, and all throughways across the valley ground to a dead halt.

Even more punishing than deadlocks was the agony of cross-guild round-trip verifications.

Even when an order avoided a deadlock, whenever components were ready for final assembly, the craftsmen were bound by the strict custom of “Two-Fold Interrogation.” Each guild locked its finished piece inside an iron strongbox and dispatched mounted couriers across the mountain trails to the central commissary: “Ward X is prepared; do we have clearance to commit?” Only after the commissary collected stamped seals from all participating guilds did he release the return tokens: “Clearance granted; unseal and combine!”

If a mounted courier slipped down a rain-soaked ravine or broke a leg along the mountain pass, the participating guilds remained paralyzed with tools suspended in midair, forbidden from moving an inch. And if a single casting cracked during final quench, all involved guilds were forced to smash their half-finished components into dust, flushing weeks of grueling labor down the drain.

The master craftsmen gathered at the valley teahouses, beating their chests in despair: “The more wondrous the artifact, the more guilds it must involve. Today, we spend less than a fraction of our hours swinging hammers or throwing shuttles; the rest is squandered guarding locked workbenches and straining our ears for the distant gallop of mountain couriers!”

The chronic dysfunction festered until a master logician and scholar from Mount Shaohua took command of the valley’s Registry Bureau.

Ascending the nine-story Celestial Bell Tower at the valley’s center, the master convened the masters of all nine guilds and proclaimed a revolutionary doctrine:

“All human affairs freeze in gridlock because men scramble for precedence; they stall in doubt because none know the progress of another. Henceforth: seal the universal sequence first, and only then strike the hammer!”

The master established the Grand Scribe’s Pavilion within the Bell Tower and instituted three immutable laws:

First, Chiming the Epoch and Total Ordering. The nine guilds dismantled their private intake desks. Every order, grand or humble, involving one guild or nine, had to be submitted directly to the Celestial Bell Tower. At the turn of every half-hour, the bronze bell atop the tower boomed across the valley (Epoch). With each chime, the scribes bundled all incoming orders received during that window and inscribed them onto a master scroll with absolute, immutable ordinal numbers: #1, #2, #3, and onward. Instantly, via a high-tension cableway network, identical copies of the “Preordained Celestial Chronicle” were zipped to the benches of all nine guilds simultaneously!

Second, Deterministic Execution and the Extinction of Deadlocks. Inside each workshop, the masters unrolled their copies of the scroll. The ironmaster examined only entries concerning steel; the woodcarver observed only timber. But regardless of the guild, every artisan was bound by law to allocate anvils, lathes, and master hands in the exact sequence of the roster numbers! If Order #7 and Order #9 both required Forge Anvil No. 1, the anvil was unconditionally assigned to Order #7 first. The craftsmen of Order #9 stood patiently with hands tucked into sleeves. Meanwhile, the contested loom in the Loom Ward was allocated to Order #7 with identical, unswerving precision. Because all nine guilds across the valley honored the exact same preordained sequence, cyclical deadlocks were mathematically annihilated before a single chisel touched wood or iron!

The guildmasters marveled at the orderliness, yet their deepest anxiety remained: “Master, establishing sequence is straightforward, but how do we coordinate across guilds? When the Ironworks finishes forging an axle, how does it know whether the Timber Ward’s carriage frame has succeeded? Must we not send couriers to inquire if they are ready to commit?”

The master logician laughed toward the heavens, unveiling the crowning jewel of the architecture:

“Third, Silent Harmony without Distributed Interrogation!”

“What is the Celestial Chronicle? The chronicle records that Order #7 is the Celestial Steed Carriage. That single truth means that every craftsman across all nine guilds already knows the complete destiny of Order #7! When the master smith finishes forging the carriage axle, he reports neither to the Bell Tower nor to the Timber Ward. He simply plunges the white-hot axle into cold mountain springs, secures it inside a sealed iron basket, hooks it to the aerial transit cable, and pulls the release lever—ziiiiip! The basket hurtles straight across the sky into the courtyard of the Timber Ward! When the timber artisan reaches Order #7 on his own copy of the chronicle, he reaches out his arm, lifts the axle from the basket, aligns the tenons, and hammers it home!”

An elderly artisan asked with a trembling voice: “Master… do we truly send no messenger to ask: ‘Are you prepared? Shall we finalize the contract?’”

“Not a single word!” the master replied, eyes shining with fierce conviction. “In former times, you interrogated one another because neither knew when the other would begin, when they would rest, or if they would abandon the work. Today, the destiny and ordering of the entire valley were ratified by heaven the moment the bell chimed! So long as each artisan executes his deterministic duty, the outcome unfolds as inevitably as the dawn. When you know with mathematical certainty that your brother artisan will deliver the axle at this very hour, what need is there to squander time whispering across the mountains?”

From that day forward, frantic couriers vanished from the mountain paths of the Valley of Heavenly Craft. Brawls over shared forges ceased forever.

Every half-hour, the bronze bell tolled, and ten thousand tools stirred in unison. Anvils rang like thunder, shuttles flew like streaks of silver light, and intricate components glided along high-tension aerial cables with breathtaking precision. The nine disparate guilds functioned as one colossal, harmonious celestial automaton, pulsing with boundless power.

Merchants from across the empire gasped in awe: complex multi-guild masterpieces that had previously dragged through weeks of bureaucratic wrangling now materialized in continuous, effortless triumph to the rhythmic tolling of the great bronze bell!

—By now you’ve probably recognized it: this revolutionary architecture of establishing global consensus on transaction ordering before locks are acquired, executing deterministic state transitions locally without distributed deadlocks, and eliminating the distributed commit protocol entirely, is the groundbreaking deterministic distributed transaction framework known as Calvin—introduced in 2012 by Daniel Abadi’s team at Yale University, which fundamentally shattered the latency and lock-contention shackles of classic Two-Phase Commit (2PC).

What it is

Calvin was introduced in 2012 by Alexander Thomson, Daniel Abadi, and their collaborators in their landmark ACM SIGMOD paper, “Calvin: Fast Distributed Transactions for Partitioned Database Systems.”

In conventional distributed relational database systems (such as those employing classical Two-Phase Commit 2PC, Google Spanner, or Google Percolator), transaction processing follows a dynamically interleaved execution model:

1. Client transactions begin executing across partition nodes, dynamically requesting and acquiring row-level locks from storage engines as read/write operations occur.
2. Once all operations finish, a designated coordinator initiates a distributed Two-Phase Commit (2PC: Prepare and Commit phases) across all participating shards.
3. Shards must hold these locks continuously throughout the 2PC protocol, spanning multiple network round-trips (RTTs) across local networks or wide-area networks (WAN), until the transaction finally commits.

This conventional paradigm incurs catastrophic performance overheads in distributed environments:

• Excessive Lock Duration (Lock Contention & Long Hold Times): Locks cover not just CPU computation time, but the unpredictable latency of network communication and distributed consensus.
• Rampant Deadlocks and Aborts: Because transactions acquire locks dynamically in unpredictable orders across shards, distributed deadlocks are common. Resolving them requires aborting and retrying transactions, causing throughput to collapse under high contention.

Calvin introduced a paradigm shift: completely decouple the consensus/sequencing layer from execution and concurrency control. Before any transaction touches a single record or acquires a single lock, agree upon a globally unique, totally ordered transaction sequence via a replicated consensus log.

Calvin’s architecture consists of three core layers:

1. Sequencing Layer (Sequencer):
   • Transactions submitted by clients are received by independent sequencer nodes.
   • Sequencers group transactions into discrete temporal batches known as epochs (typically 10 milliseconds). Within each epoch, each transaction is assigned a globally unique, monotonically increasing sequence number.
   • The sequencers run a consensus protocol (such as Paxos or Raft) to replicate the epoch batches across all replica sets, guaranteeing that every storage node receives an identical, immutable, and globally ordered transaction sequence log.
2. Scheduling and Concurrency Control Layer (Deterministic Locking):
   • When a storage shard receives an epoch batch, its local deterministic scheduler allocates locks.
   • Deadlock-Free Deterministic Locking: The scheduler requests locks strictly in the ascending order of the global transaction sequence. Because every node requests locks in the exact same predefined sequence, circular wait conditions are mathematically impossible. Calvin requires no deadlock detection graphs, no wound-wait timeouts, and no transaction aborts due to lock contention!
3. Execution Layer without Two-Phase Commit (No 2PC & Point-to-Point Messaging):
   • For distributed transactions spanning multiple partitions, Calvin analyzes read and write sets to classify nodes as active participants (nodes performing writes or executing logic) or passive participants (nodes merely providing read-only records).
   • Passive participants read requested records locally in sequence order and forward the data point-to-point directly to the active participants via network pipes.
   • Active participants assemble the remote read records, deterministically execute the transaction logic, and write modifications to storage.
   • Complete Elimination of Two-Phase Commit (Elimination of 2PC): Because the global order was finalized prior to execution and transaction logic is strictly deterministic, every participating node knows with certainty that all other participants will execute successfully. Nodes never exchange Prepare or Commit rounds! Locks are held solely for the microseconds required for local memory/disk computation, releasing immediately upon completion without waiting for WAN acknowledgments.

For transactions where read and write sets cannot be statically determined upfront (e.g., secondary index lookups or pointer chasing where write targets depend on read values), Calvin provides an Optimistic Lock-free Reconnaissance Phase (OLR): an initial lightweight, non-transactional read query inspects the database to predict key targets. The fully specified transaction is then submitted to the sequencer. If the underlying data mutated between reconnaissance and execution, the transaction aborts deterministically at the sequencer level and retries.

Why it matters

Calvin fundamentally reshaped distributed database theory and modern high-throughput transactional architectures:

1. Eliminating the Network-Bound Lock Bottleneck: Classic 2PC inherently binds lock hold times to network latency. By ordering transactions before locking, Calvin isolates network consensus to the sequencing phase, allowing locks to be held strictly during local execution. Distributed transaction throughput approaches that of single-node engines.
2. Resilience under Extreme Contention and Geo-Replication: In multi-region active-active deployments (where WAN latency exceeds 50–100ms) or hot-spot workloads (e.g., flash sales, financial ledgers), conventional 2PC throughput collapses due to lock queuing and abort cascades. Calvin maintains consistent, predictable high throughput regardless of network distance.
3. Foundation of Modern NewSQL and Deterministic Engines: Calvin directly inspired production-grade cloud databases such as FaunaDB (whose transaction engine is a commercialized implementation of Calvin), as well as hybrid deterministic systems like SLOG and BOHM. Understanding Calvin is essential for mastering how modern distributed systems transcend the boundaries of classical Two-Phase Commit.

Metaphor mapping

• Nine artisan guilds operating across the valley → Distributed database partitions and storage shards (Database Partitions / Shards)
• Apprentices brawling over shared workbenches and halting work for days → Distributed deadlocks caused by unordered lock acquisition (Distributed Deadlocks)
• Locking components inside strongboxes while couriers verify clearance across guilds → Classic Two-Phase Commit (2PC) holding locks across network round-trips (Distributed 2PC Lock Holding)
• Mountain landslides stranding couriers and freezing workshops → 2PC coordinator failure or network latency causing blocking (2PC Blocking Problem)
• Central Celestial Bell Tower and Grand Scribe’s Pavilion → Independent consensus and sequencing layer (Sequencer Layer / Consensus Log)
• Chiming the bell every half-hour to seal all orders into a scroll → Batching transactions into discrete epochs (Epoch Batching)
• Stamping identical sequential numbers and broadcasting scrolls via cableway → Replicating a globally unique, totally ordered sequence log via Paxos/Raft (Replicated Global Total Order)
• Allocating anvils and lathes in strict scroll sequence without dispute → Deterministic locking acquiring locks in global sequence order without deadlocks (Deterministic Locking / Deadlock-Free Scheduler)
• Sending forged axle directly via aerial basket to the Timber Ward → Active participants receiving read records point-to-point from passive participants (Point-to-Point Data Forwarding)
• Timber Ward assembling the carriage without sending couriers to verify commit status → Deterministic execution completely eliminating Two-Phase Commit (Elimination of Distributed 2PC)
• Sketching blueprints to deduce which workshops an order will touch → Optimistic reconnaissance phase predicting dynamic read/write sets (Optimistic Reconnaissance Phase)