Shrinking Phase — Two-Phase Locking Protocol | OneNoughtOne

0/105

Database Management SystemsTwo-Phase Locking (2PL)

Two-Phase Locking Protocol

LevelIntermediate

Duration60 mins

TopicTwo-Phase Locking (2PL)

3 / 5

Shrinking Phase

The Art of Letting Go

If the Growing Phase is about establishing territory, the Shrinking Phase is about gracefully relinquishing control. This phase determines when locked resources become available to other transactions, directly impacting system concurrency and throughput. The Shrinking Phase might seem straightforward—just release the locks when you're done, right? But the timing and strategy of lock releases have profound implications for system performance, recoverability, and even correctness. A single premature release can cascade into data corruption. Holding locks too long starves other transactions of resources. In this page, we'll explore every aspect of the Shrinking Phase, from the fundamental mechanics of lock release to the sophisticated strategies that production databases employ to balance correctness with concurrency.

What You Will Learn

By the end of this page, you will understand the mechanics of lock release during the shrinking phase, how lock release timing affects other waiting transactions, the relationship between lock release and transaction commitment, and why premature release can lead to serious correctness problems.

Mechanics of Lock Release

During the Shrinking Phase, a transaction methodically releases the locks it acquired during the Growing Phase. While conceptually simple, the actual process involves careful coordination with the lock manager and waiting transactions. The Lock Release Lifecycle:

Converting Mermaid diagram...

Step 1: Release Request Issuance When a transaction decides to release a lock (either explicitly or implicitly at COMMIT/ROLLBACK), it signals the lock manager. The release request specifies: Transaction ID (TID): The transaction releasing the lock Data Item Identifier: The specific item being unlocked Lock Mode: The mode being released (relevant for transactions holding multiple lock types) Step 2: Lock Table Update The lock manager updates its internal data structures:

lock_release_algorithm.pseudocode
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
FUNCTION process_lock_release(tid, data_item):
    // Validate the release request
    current_locks = lock_table.get(data_item)
    
    IF tid NOT IN current_locks.holders:
        RAISE ERROR "Transaction does not hold lock on item"
    
    // Remove transaction from holders list
    current_locks.holders.remove(tid)
    lock_table.update(data_item, current_locks)
    
    // Log the release for recovery purposes
    log_record = create_log("UNLOCK", tid, data_item), 
    write_to_log(log_record)
    
    // Check if waiting transactions can now proceed
    grant_waiting_requests(data_item)
    
    RETURN SUCCESS
 
 
FUNCTION grant_waiting_requests(data_item):
    wait_queue = get_wait_queue(data_item)
    current_holders = lock_table.get(data_item).holders
    
    // Process wait queue in order
    granted_list = []
    
    FOR EACH (waiting_tid, requested_mode) IN wait_queue:
        // Check if request is now compatible with current holders
        IF is_compatible_with_all(requested_mode, current_holders):
            // Grant the lock
            lock_table.add_holder(data_item, waiting_tid, requested_mode)
            wait_queue.remove(waiting_tid)
            granted_list.append(waiting_tid)
            
            // Update current holders for next iteration
            current_holders.append((waiting_tid, requested_mode))
        ELSE:
            // Cannot grant - stop processing (preserve FIFO for writes)
            // Continue checking for compatible shared requests (optional)
            IF requested_mode == EXCLUSIVE:
                BREAK  // Don't skip over exclusive requests
    
    // Wake up all granted transactions
    FOR EACH tid IN granted_list:
        signal_transaction(tid, LOCK_GRANTED)

The Cascading Effect

A single lock release can trigger a cascade of grants. If Transaction A releases an exclusive lock, and the wait queue contains [B(Shared), C(Shared), D(Exclusive), E(Shared)], the lock manager grants locks to B, C simultaneously. D still waits (conflicts with B, C), and E might be granted (compatible with B, C) or wait (if strict FIFO is enforced).

When Does the Shrinking Phase Begin?

The transition from Growing Phase to Shrinking Phase is a critical moment in a transaction's lifecycle. This transition occurs the instant a transaction releases its first lock—and once crossed, there's no going back. The Lock Point Revisited:

Transaction Phase Transition
Time	Operation	Phase	Lock Count
t₁	BEGIN TRANSACTION	—	0
t₂	Lock(A)	Growing	1
t₃	Read(A)	Growing	1
t₄	Lock(B)	Growing	2
t₅	Write(B)	Growing	2
t₆	Lock(C)	Growing	3
t₇	Read(C)	Growing	3
t₈	Unlock(A) ⚡	→ Shrinking	2
t₉	Unlock(B)	Shrinking	1
t₁₀	Unlock(C)	Shrinking	0
t₁₁	COMMIT	Complete	0

Key Observations: Lock Point occurs at t₇ — This is the moment of maximum lock holding. All locks the transaction will ever need have been acquired. Phase transition occurs at t₈ — The first Unlock operation marks the beginning of the Shrinking Phase. This is an irreversible state change. No new locks after t₈ — Once Unlock(A) occurs, the transaction cannot acquire Lock(D) or any other new lock. Attempting to do so violates 2PL. Work can continue — Notice that reads and writes can occur in both phases. The phase distinction only constrains lock operations, not data operations.

The Point of No Return

In basic 2PL, the first unlock is irrevocable. Even if the transaction later realizes it needs to access another data item, it cannot acquire a new lock. This constraint is why transaction designers must carefully plan which data items will be needed before releasing any locks. In practice, this is why Strict 2PL (holding all locks until commit) became the dominant variant.

2pl_violation_example.sql
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
-- INCORRECT: This transaction violates 2PL
 
BEGIN TRANSACTION;
 
Lock(Account_A);        -- Growing phase
Read(Account_A);        -- balance = $1000
Unlock(Account_A);      -- Shrinking phase begins ⚡
 
-- Transaction decides it needs to check Account_B
-- Perhaps based on a business rule discovered after reading A
 
Lock(Account_B);        -- ❌ VIOLATION! Cannot acquire after releasing
                        -- This would re-enter growing phase
                        -- 2PL FORBIDS THIS
 
-- The transaction is now INVALID under 2PL
-- Database would either:
--   (a) ABORT the transaction
--   (b) Block and wait (if using Conservative 2PL variant)
--   (c) Never happen (if using Strict 2PL - locks held until commit)

Lock Release Ordering

A subtle but important aspect of the Shrinking Phase is the order in which locks are released. While 2PL correctness doesn't depend on release order (any order maintains serializability), the choice of ordering significantly impacts concurrency and deadlock behavior.

LIFO Release (Stack Order)

•Pattern: Release in reverse acquisition order
•Example: Lock(A), Lock(B), Lock(C) → Unlock(C), Unlock(B), Unlock(A)
•Advantage: Matches natural nested access patterns
•Advantage: Can help prevent certain deadlocks
•Use Case: Hierarchical locking scenarios

FIFO Release (Queue Order)

•Pattern: Release in acquisition order
•Example: Lock(A), Lock(B), Lock(C) → Unlock(A), Unlock(B), Unlock(C)
•Advantage: Releases "oldest" locks earliest
•Advantage: May improve concurrency for hot resources
•Use Case: Sequential processing pipelines

Strategic Release for Concurrency: In practice, the optimal release order often depends on workload patterns:

strategic_release.sql
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
-- Scenario: Processing an order (involves inventory, account, shipping)
 
BEGIN TRANSACTION;
 
-- Growing Phase
Lock(Inventory_Item_42);     -- Check stock
Lock(Customer_Account_7);    -- Check payment method
Lock(Shipping_Zone_3);       -- Calculate shipping
 
-- Process the order...
Read(Inventory_Item_42);     -- 50 units available
Read(Customer_Account_7);    -- Credit card on file
Read(Shipping_Zone_3);       -- $5.99 shipping rate
 
-- Shrinking Phase - Strategic Release Order
 
-- OPTION A: Release hottest resource first
Unlock(Inventory_Item_42);   -- High contention - release early
Unlock(Customer_Account_7);  -- Medium contention
Unlock(Shipping_Zone_3);     -- Low contention - release last
 
-- OPTION B: Release least-needed first  
Unlock(Shipping_Zone_3);     -- Done with shipping calc
Unlock(Customer_Account_7);  -- Done with payment check
Unlock(Inventory_Item_42);   -- Keep stock locked longest
 
-- The 2PL guarantee holds regardless of order
-- But system throughput may vary based on contention patterns

Practical Guidance

In high-contention systems, consider releasing the most contended locks first during the shrinking phase. This unblocks waiting transactions sooner, improving overall system throughput. However, ensure you truly won't need those resources again—remember, you can't re-acquire after releasing!

The Danger of Early Release

We've emphasized that 2PL guarantees serializability. But let's deeply understand why early release violates this guarantee. This understanding is crucial for debugging concurrency anomalies.

early_release_corruption.sql
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
-- THE LOST UPDATE VIA EARLY RELEASE
 
-- Scenario: Two transactions updating the same account balance
-- Both want to add to the balance based on reading the current value
 
-- Initial state: Account_A.balance = $1000
 
-- Transaction T₁ (Early Release - INCORRECT)
T₁: Lock(Account_A)           
T₁: Read(Account_A) → $1000   
T₁: Unlock(Account_A)         -- ❌ EARLY RELEASE!
T₁: compute: new_balance = 1000 + 200 = $1200
-- T₁ needs to write... but released the lock!
 
-- Transaction T₂ (executes during T₁'s computation)
T₂: Lock(Account_A)           -- Granted! T₁ released it
T₂: Read(Account_A) → $1000   -- Still $1000
T₂: compute: new_balance = 1000 + 300 = $1300
T₂: Write(Account_A) ← $1300  -- Balance now $1300
T₂: Unlock(Account_A)
T₂: COMMIT
 
-- T₁ continues (assumes it still has valid data)
T₁: Lock(Account_A)           -- Re-acquired (VIOLATES 2PL)
T₁: Write(Account_A) ← $1200  -- OVERWRITES T₂'s update!
T₁: Unlock(Account_A)
T₁: COMMIT
 
-- RESULT: balance = $1200
-- CORRECT: balance should be $1000 + $200 + $300 = $1500
-- LOST: T₂'s $300 deposit vanished!

The Lost Update Problem

This is a classic lost update scenario. By releasing the lock early, T₁ allowed T₂ to read stale data and base its update on it. Then T₁'s re-acquisition (which violates 2PL) overwrote T₂'s committed work. If T₁ had held the lock throughout its growing phase, T₂ would have waited, read $1200, and written $1500.

Why 2PL Prevents This: Under proper 2PL, T₁ would: Lock Account_A Read \to $1000 Compute new value Write \leftarrow $1200 Only then Unlock(Account_A) T₂ would block at step 1, waiting for T₁'s exclusive lock. After T₁ commits: T₂ Lock(Account_A) — granted T₂ Read \to $1200 (sees T₁'s update) T₂ Compute: 1200 + 300 = $1500 T₂ Write \leftarrow $1500 T₂ Unlock and Commit Result: $1500 — Both updates are correctly applied.

Consequences of Early Release

•Lost Updates — Concurrent writes overwrite each other's changes
•Inconsistent Reads — Transactions observe impossible database states
•Constraint Violations — Invariants broken by interleaved operations
•Cascading Failures — One corrupted value propagates to dependent computations
•Audit Trail Corruption — Historical records become unreliable

Implicit Release at Commit and Abort

In most database systems, explicit lock release commands are rare. Instead, locks are released implicitly as part of transaction termination (COMMIT or ROLLBACK). This approach simplifies programming and, crucially, enables Strict 2PL . The Commit-Release Pattern:

implicit_release.sql
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
-- Typical transaction - no explicit lock/unlock statements
 
BEGIN TRANSACTION;
 
-- Implicit Lock(customers) on touched rows
UPDATE customers SET last_purchase = NOW() WHERE id = 42;
 
-- Implicit Lock(orders) on new row
INSERT INTO orders (customer_id, total) VALUES (42, 99.99);
 
-- Implicit Lock(inventory) on touched rows  
UPDATE inventory SET quantity = quantity - 1 WHERE product_id = 101;
 
COMMIT;  -- ← ALL locks released here atomically
 
-- What happens at COMMIT:
-- 1. Transaction log is flushed (durability)
-- 2. Commit record written to log
-- 3. Lock Manager notified of transaction completion
-- 4. ALL locks held by this transaction released
-- 5. Waiting transactions notified
-- 6. Transaction resources freed

Abort and Rollback Release: When a transaction aborts (either explicitly via ROLLBACK or due to an error/deadlock), locks are also released—but with additional considerations:

Lock Release: Commit vs Abort
Aspect	COMMIT	ABORT/ROLLBACK
Changes Persisted	Yes	No — Changes undone
Lock Release Timing	After commit record durable	After undo complete
Waiting Txns See	Committed values	Original values (pre-abort)
Log Records	Commit record written	Abort record + CLRs written
Resource Cleanup	Standard cleanup	Buffer pool pages invalidated

Why Wait for Undo Before Release?

When a transaction aborts, its changes must be undone before locks are released. If locks were released while undo was still in progress, other transactions might read partially-undone data. This would violate atomicity—transactions would observe states that never "existed" in any committed view of the database.

Shrinking Phase and System Concurrency

The Shrinking Phase directly controls how quickly resources return to the shared pool. Understanding this relationship is essential for capacity planning and performance tuning. The Concurrency Equation: Consider a simplified model where: N transactions want to access resource R Each transaction holds R for time T Only one transaction can hold the exclusive lock at a time Maximum Throughput = 1/T transactions per time unit The Shrinking Phase duration directly impacts T. Longer lock holding = Lower throughput.

concurrency_analysis.txt
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
SCENARIO: E-commerce checkout (exclusive lock on inventory item)
 
Processing Timeline for Single Item:
 
[────────────── Transaction Duration ──────────────]
[──Growing──][────────Work────────][──Shrinking──]
   Lock(I)     Process     Write     Unlock(I)  COMMIT
   
Timeline with Basic 2PL (early shrinking allowed):
T₁: |--Lock--|---Work---|---Shrink/Commit---|
T₂:                     |--Lock--|---Work---|---Shrink/Commit---|
T₃:                                         |--Lock--|---Work---|
 
Total time for 3 transactions: ~2.5x single transaction
(Overlap possible in shrinking phase)
 
 
Timeline with Strict 2PL (shrinking only at commit):
T₁: |--Lock--|---Work---|-Commit-|
T₂:                              |--Lock--|---Work---|-Commit-|
T₃:                                                           |--Lock--...
 
Total time for 3 transactions: 3x single transaction
(No overlap - each waits for previous to fully complete)
 
 
TRADE-OFF ANALYSIS:
┌──────────────────┬────────────────┬────────────────────┐
│ Property         │ Basic 2PL      │ Strict 2PL         │
├──────────────────┼────────────────┼────────────────────┤
│ Concurrency      │ Higher         │ Lower              │
│ Recoverability   │ Not guaranteed │ Guaranteed         │
│ Cascading Abort  │ Possible       │ Impossible         │
│ Lock Hold Time   │ Shorter        │ Longer             │
│ Deadlock Risk    │ Similar        │ Similar            │
└──────────────────┴────────────────┴────────────────────┘

The Recoverability Trade-off: Basic 2PL allows lock release before commit, which enables higher concurrency but introduces a subtle problem: cascading aborts . If T₁ releases a lock on X before committing, and T₂ reads X (with the value T₁ wrote), then T₂ depends on T₁'s success. If T₁ later aborts, T₂ has read data that was never actually committed—T₂ must also be aborted! This can cascade through many transactions. Strict 2PL solves this by holding all locks until commit. Since no other transaction can read uncommitted data, aborts never cascade. The trade-off is reduced concurrency.

Production Reality

Virtually all production database systems implement Strict 2PL (or its stronger variant, Rigorous 2PL) precisely because of the recoverability guarantee. The concurrency cost is considered acceptable compared to the complexity of handling cascading aborts. We'll explore these variants in detail in the next module.

Lock Release in Distributed Systems

When databases span multiple nodes, the Shrinking Phase becomes significantly more complex. Lock release must be coordinated across the distributed system to maintain correctness. The Distributed Commit Challenge:

Converting Mermaid diagram...

Two-Phase Commit (2PC) and Lock Release: In distributed transactions using Two-Phase Commit (2PC), the Shrinking Phase is intimately tied to the commit protocol: Prepare Phase: Each node continues holding its locks. No releases occur. Vote: Nodes vote YES (can commit) or NO (must abort). Locks still held. Commit Decision: Coordinator decides COMMIT or ABORT. Commit Phase: Each node commits/aborts and only then releases locks. Critical Constraint: All nodes must maintain their locks through the entire prepare phase. If a node released locks after voting YES but before receiving the final COMMIT, another transaction could see incomplete state.

Distributed Lock Release Challenges

•Lock Holding During Prepare: Nodes hold locks during the uncertainty window between PREPARE and COMMIT/ABORT decision
•Coordinator Failure: If coordinator fails after PREPARE, nodes are stuck holding locks until recovery
•Network Partitions: Nodes may be unable to receive the final decision, blocking lock release
•Clock Skew: Distributed timeout detection is complicated by clock differences
•Ordered Release: May need to release locks in specific order across nodes to prevent distributed deadlocks

Modern Approaches

Modern distributed databases like CockroachDB, Spanner, and YugabyteDB use sophisticated variations of 2PC with features like parallel commits, pipelining, and optimistic lock release to minimize the performance impact of distributed lock coordination. These systems maintain 2PL semantics while achieving high throughput across distributed nodes.

Database Management SystemsTwo-Phase Locking (2PL)

Two-Phase Locking Protocol

LevelIntermediate

Duration60 mins

TopicTwo-Phase Locking (2PL)

3 / 5

Shrinking Phase

The Art of Letting Go

If the Growing Phase is about establishing territory, the Shrinking Phase is about gracefully relinquishing control. This phase determines when locked resources become available to other transactions, directly impacting system concurrency and throughput. The Shrinking Phase might seem straightforward—just release the locks when you're done, right? But the timing and strategy of lock releases have profound implications for system performance, recoverability, and even correctness. A single premature release can cascade into data corruption. Holding locks too long starves other transactions of resources. In this page, we'll explore every aspect of the Shrinking Phase, from the fundamental mechanics of lock release to the sophisticated strategies that production databases employ to balance correctness with concurrency.

What You Will Learn

By the end of this page, you will understand the mechanics of lock release during the shrinking phase, how lock release timing affects other waiting transactions, the relationship between lock release and transaction commitment, and why premature release can lead to serious correctness problems.

Mechanics of Lock Release

During the Shrinking Phase, a transaction methodically releases the locks it acquired during the Growing Phase. While conceptually simple, the actual process involves careful coordination with the lock manager and waiting transactions. The Lock Release Lifecycle:

Converting Mermaid diagram...

Step 1: Release Request Issuance When a transaction decides to release a lock (either explicitly or implicitly at COMMIT/ROLLBACK), it signals the lock manager. The release request specifies: Transaction ID (TID): The transaction releasing the lock Data Item Identifier: The specific item being unlocked Lock Mode: The mode being released (relevant for transactions holding multiple lock types) Step 2: Lock Table Update The lock manager updates its internal data structures:

lock_release_algorithm.pseudocode
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
FUNCTION process_lock_release(tid, data_item):
    // Validate the release request
    current_locks = lock_table.get(data_item)
    
    IF tid NOT IN current_locks.holders:
        RAISE ERROR "Transaction does not hold lock on item"
    
    // Remove transaction from holders list
    current_locks.holders.remove(tid)
    lock_table.update(data_item, current_locks)
    
    // Log the release for recovery purposes
    log_record = create_log("UNLOCK", tid, data_item), 
    write_to_log(log_record)
    
    // Check if waiting transactions can now proceed
    grant_waiting_requests(data_item)
    
    RETURN SUCCESS
 
 
FUNCTION grant_waiting_requests(data_item):
    wait_queue = get_wait_queue(data_item)
    current_holders = lock_table.get(data_item).holders
    
    // Process wait queue in order
    granted_list = []
    
    FOR EACH (waiting_tid, requested_mode) IN wait_queue:
        // Check if request is now compatible with current holders
        IF is_compatible_with_all(requested_mode, current_holders):
            // Grant the lock
            lock_table.add_holder(data_item, waiting_tid, requested_mode)
            wait_queue.remove(waiting_tid)
            granted_list.append(waiting_tid)
            
            // Update current holders for next iteration
            current_holders.append((waiting_tid, requested_mode))
        ELSE:
            // Cannot grant - stop processing (preserve FIFO for writes)
            // Continue checking for compatible shared requests (optional)
            IF requested_mode == EXCLUSIVE:
                BREAK  // Don't skip over exclusive requests
    
    // Wake up all granted transactions
    FOR EACH tid IN granted_list:
        signal_transaction(tid, LOCK_GRANTED)

The Cascading Effect

A single lock release can trigger a cascade of grants. If Transaction A releases an exclusive lock, and the wait queue contains [B(Shared), C(Shared), D(Exclusive), E(Shared)], the lock manager grants locks to B, C simultaneously. D still waits (conflicts with B, C), and E might be granted (compatible with B, C) or wait (if strict FIFO is enforced).

When Does the Shrinking Phase Begin?

The transition from Growing Phase to Shrinking Phase is a critical moment in a transaction's lifecycle. This transition occurs the instant a transaction releases its first lock—and once crossed, there's no going back. The Lock Point Revisited:

Transaction Phase Transition
Time	Operation	Phase	Lock Count
t₁	BEGIN TRANSACTION	—	0
t₂	Lock(A)	Growing	1
t₃	Read(A)	Growing	1
t₄	Lock(B)	Growing	2
t₅	Write(B)	Growing	2
t₆	Lock(C)	Growing	3
t₇	Read(C)	Growing	3
t₈	Unlock(A) ⚡	→ Shrinking	2
t₉	Unlock(B)	Shrinking	1
t₁₀	Unlock(C)	Shrinking	0
t₁₁	COMMIT	Complete	0

Key Observations: Lock Point occurs at t₇ — This is the moment of maximum lock holding. All locks the transaction will ever need have been acquired. Phase transition occurs at t₈ — The first Unlock operation marks the beginning of the Shrinking Phase. This is an irreversible state change. No new locks after t₈ — Once Unlock(A) occurs, the transaction cannot acquire Lock(D) or any other new lock. Attempting to do so violates 2PL. Work can continue — Notice that reads and writes can occur in both phases. The phase distinction only constrains lock operations, not data operations.

The Point of No Return

In basic 2PL, the first unlock is irrevocable. Even if the transaction later realizes it needs to access another data item, it cannot acquire a new lock. This constraint is why transaction designers must carefully plan which data items will be needed before releasing any locks. In practice, this is why Strict 2PL (holding all locks until commit) became the dominant variant.

2pl_violation_example.sql
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
-- INCORRECT: This transaction violates 2PL
 
BEGIN TRANSACTION;
 
Lock(Account_A);        -- Growing phase
Read(Account_A);        -- balance = $1000
Unlock(Account_A);      -- Shrinking phase begins ⚡
 
-- Transaction decides it needs to check Account_B
-- Perhaps based on a business rule discovered after reading A
 
Lock(Account_B);        -- ❌ VIOLATION! Cannot acquire after releasing
                        -- This would re-enter growing phase
                        -- 2PL FORBIDS THIS
 
-- The transaction is now INVALID under 2PL
-- Database would either:
--   (a) ABORT the transaction
--   (b) Block and wait (if using Conservative 2PL variant)
--   (c) Never happen (if using Strict 2PL - locks held until commit)

Lock Release Ordering

A subtle but important aspect of the Shrinking Phase is the order in which locks are released. While 2PL correctness doesn't depend on release order (any order maintains serializability), the choice of ordering significantly impacts concurrency and deadlock behavior.

LIFO Release (Stack Order)

•Pattern: Release in reverse acquisition order
•Example: Lock(A), Lock(B), Lock(C) → Unlock(C), Unlock(B), Unlock(A)
•Advantage: Matches natural nested access patterns
•Advantage: Can help prevent certain deadlocks
•Use Case: Hierarchical locking scenarios

FIFO Release (Queue Order)

•Pattern: Release in acquisition order
•Example: Lock(A), Lock(B), Lock(C) → Unlock(A), Unlock(B), Unlock(C)
•Advantage: Releases "oldest" locks earliest
•Advantage: May improve concurrency for hot resources
•Use Case: Sequential processing pipelines

Strategic Release for Concurrency: In practice, the optimal release order often depends on workload patterns:

strategic_release.sql
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
-- Scenario: Processing an order (involves inventory, account, shipping)
 
BEGIN TRANSACTION;
 
-- Growing Phase
Lock(Inventory_Item_42);     -- Check stock
Lock(Customer_Account_7);    -- Check payment method
Lock(Shipping_Zone_3);       -- Calculate shipping
 
-- Process the order...
Read(Inventory_Item_42);     -- 50 units available
Read(Customer_Account_7);    -- Credit card on file
Read(Shipping_Zone_3);       -- $5.99 shipping rate
 
-- Shrinking Phase - Strategic Release Order
 
-- OPTION A: Release hottest resource first
Unlock(Inventory_Item_42);   -- High contention - release early
Unlock(Customer_Account_7);  -- Medium contention
Unlock(Shipping_Zone_3);     -- Low contention - release last
 
-- OPTION B: Release least-needed first  
Unlock(Shipping_Zone_3);     -- Done with shipping calc
Unlock(Customer_Account_7);  -- Done with payment check
Unlock(Inventory_Item_42);   -- Keep stock locked longest
 
-- The 2PL guarantee holds regardless of order
-- But system throughput may vary based on contention patterns

Practical Guidance

In high-contention systems, consider releasing the most contended locks first during the shrinking phase. This unblocks waiting transactions sooner, improving overall system throughput. However, ensure you truly won't need those resources again—remember, you can't re-acquire after releasing!

The Danger of Early Release

We've emphasized that 2PL guarantees serializability. But let's deeply understand why early release violates this guarantee. This understanding is crucial for debugging concurrency anomalies.

early_release_corruption.sql
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
-- THE LOST UPDATE VIA EARLY RELEASE
 
-- Scenario: Two transactions updating the same account balance
-- Both want to add to the balance based on reading the current value
 
-- Initial state: Account_A.balance = $1000
 
-- Transaction T₁ (Early Release - INCORRECT)
T₁: Lock(Account_A)           
T₁: Read(Account_A) → $1000   
T₁: Unlock(Account_A)         -- ❌ EARLY RELEASE!
T₁: compute: new_balance = 1000 + 200 = $1200
-- T₁ needs to write... but released the lock!
 
-- Transaction T₂ (executes during T₁'s computation)
T₂: Lock(Account_A)           -- Granted! T₁ released it
T₂: Read(Account_A) → $1000   -- Still $1000
T₂: compute: new_balance = 1000 + 300 = $1300
T₂: Write(Account_A) ← $1300  -- Balance now $1300
T₂: Unlock(Account_A)
T₂: COMMIT
 
-- T₁ continues (assumes it still has valid data)
T₁: Lock(Account_A)           -- Re-acquired (VIOLATES 2PL)
T₁: Write(Account_A) ← $1200  -- OVERWRITES T₂'s update!
T₁: Unlock(Account_A)
T₁: COMMIT
 
-- RESULT: balance = $1200
-- CORRECT: balance should be $1000 + $200 + $300 = $1500
-- LOST: T₂'s $300 deposit vanished!

The Lost Update Problem

This is a classic lost update scenario. By releasing the lock early, T₁ allowed T₂ to read stale data and base its update on it. Then T₁'s re-acquisition (which violates 2PL) overwrote T₂'s committed work. If T₁ had held the lock throughout its growing phase, T₂ would have waited, read $1200, and written $1500.

Why 2PL Prevents This: Under proper 2PL, T₁ would: Lock Account_A Read \to $1000 Compute new value Write \leftarrow $1200 Only then Unlock(Account_A) T₂ would block at step 1, waiting for T₁'s exclusive lock. After T₁ commits: T₂ Lock(Account_A) — granted T₂ Read \to $1200 (sees T₁'s update) T₂ Compute: 1200 + 300 = $1500 T₂ Write \leftarrow $1500 T₂ Unlock and Commit Result: $1500 — Both updates are correctly applied.

Consequences of Early Release

•Lost Updates — Concurrent writes overwrite each other's changes
•Inconsistent Reads — Transactions observe impossible database states
•Constraint Violations — Invariants broken by interleaved operations
•Cascading Failures — One corrupted value propagates to dependent computations
•Audit Trail Corruption — Historical records become unreliable

Implicit Release at Commit and Abort

In most database systems, explicit lock release commands are rare. Instead, locks are released implicitly as part of transaction termination (COMMIT or ROLLBACK). This approach simplifies programming and, crucially, enables Strict 2PL . The Commit-Release Pattern:

implicit_release.sql
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
-- Typical transaction - no explicit lock/unlock statements
 
BEGIN TRANSACTION;
 
-- Implicit Lock(customers) on touched rows
UPDATE customers SET last_purchase = NOW() WHERE id = 42;
 
-- Implicit Lock(orders) on new row
INSERT INTO orders (customer_id, total) VALUES (42, 99.99);
 
-- Implicit Lock(inventory) on touched rows  
UPDATE inventory SET quantity = quantity - 1 WHERE product_id = 101;
 
COMMIT;  -- ← ALL locks released here atomically
 
-- What happens at COMMIT:
-- 1. Transaction log is flushed (durability)
-- 2. Commit record written to log
-- 3. Lock Manager notified of transaction completion
-- 4. ALL locks held by this transaction released
-- 5. Waiting transactions notified
-- 6. Transaction resources freed

Abort and Rollback Release: When a transaction aborts (either explicitly via ROLLBACK or due to an error/deadlock), locks are also released—but with additional considerations:

Lock Release: Commit vs Abort
Aspect	COMMIT	ABORT/ROLLBACK
Changes Persisted	Yes	No — Changes undone
Lock Release Timing	After commit record durable	After undo complete
Waiting Txns See	Committed values	Original values (pre-abort)
Log Records	Commit record written	Abort record + CLRs written
Resource Cleanup	Standard cleanup	Buffer pool pages invalidated

Why Wait for Undo Before Release?

When a transaction aborts, its changes must be undone before locks are released. If locks were released while undo was still in progress, other transactions might read partially-undone data. This would violate atomicity—transactions would observe states that never "existed" in any committed view of the database.

Shrinking Phase and System Concurrency

The Shrinking Phase directly controls how quickly resources return to the shared pool. Understanding this relationship is essential for capacity planning and performance tuning. The Concurrency Equation: Consider a simplified model where: N transactions want to access resource R Each transaction holds R for time T Only one transaction can hold the exclusive lock at a time Maximum Throughput = 1/T transactions per time unit The Shrinking Phase duration directly impacts T. Longer lock holding = Lower throughput.

concurrency_analysis.txt
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
SCENARIO: E-commerce checkout (exclusive lock on inventory item)
 
Processing Timeline for Single Item:
 
[────────────── Transaction Duration ──────────────]
[──Growing──][────────Work────────][──Shrinking──]
   Lock(I)     Process     Write     Unlock(I)  COMMIT
   
Timeline with Basic 2PL (early shrinking allowed):
T₁: |--Lock--|---Work---|---Shrink/Commit---|
T₂:                     |--Lock--|---Work---|---Shrink/Commit---|
T₃:                                         |--Lock--|---Work---|
 
Total time for 3 transactions: ~2.5x single transaction
(Overlap possible in shrinking phase)
 
 
Timeline with Strict 2PL (shrinking only at commit):
T₁: |--Lock--|---Work---|-Commit-|
T₂:                              |--Lock--|---Work---|-Commit-|
T₃:                                                           |--Lock--...
 
Total time for 3 transactions: 3x single transaction
(No overlap - each waits for previous to fully complete)
 
 
TRADE-OFF ANALYSIS:
┌──────────────────┬────────────────┬────────────────────┐
│ Property         │ Basic 2PL      │ Strict 2PL         │
├──────────────────┼────────────────┼────────────────────┤
│ Concurrency      │ Higher         │ Lower              │
│ Recoverability   │ Not guaranteed │ Guaranteed         │
│ Cascading Abort  │ Possible       │ Impossible         │
│ Lock Hold Time   │ Shorter        │ Longer             │
│ Deadlock Risk    │ Similar        │ Similar            │
└──────────────────┴────────────────┴────────────────────┘

The Recoverability Trade-off: Basic 2PL allows lock release before commit, which enables higher concurrency but introduces a subtle problem: cascading aborts . If T₁ releases a lock on X before committing, and T₂ reads X (with the value T₁ wrote), then T₂ depends on T₁'s success. If T₁ later aborts, T₂ has read data that was never actually committed—T₂ must also be aborted! This can cascade through many transactions. Strict 2PL solves this by holding all locks until commit. Since no other transaction can read uncommitted data, aborts never cascade. The trade-off is reduced concurrency.

Production Reality

Virtually all production database systems implement Strict 2PL (or its stronger variant, Rigorous 2PL) precisely because of the recoverability guarantee. The concurrency cost is considered acceptable compared to the complexity of handling cascading aborts. We'll explore these variants in detail in the next module.

Lock Release in Distributed Systems

When databases span multiple nodes, the Shrinking Phase becomes significantly more complex. Lock release must be coordinated across the distributed system to maintain correctness. The Distributed Commit Challenge:

Converting Mermaid diagram...

Two-Phase Commit (2PC) and Lock Release: In distributed transactions using Two-Phase Commit (2PC), the Shrinking Phase is intimately tied to the commit protocol: Prepare Phase: Each node continues holding its locks. No releases occur. Vote: Nodes vote YES (can commit) or NO (must abort). Locks still held. Commit Decision: Coordinator decides COMMIT or ABORT. Commit Phase: Each node commits/aborts and only then releases locks. Critical Constraint: All nodes must maintain their locks through the entire prepare phase. If a node released locks after voting YES but before receiving the final COMMIT, another transaction could see incomplete state.

Distributed Lock Release Challenges

•Lock Holding During Prepare: Nodes hold locks during the uncertainty window between PREPARE and COMMIT/ABORT decision
•Coordinator Failure: If coordinator fails after PREPARE, nodes are stuck holding locks until recovery
•Network Partitions: Nodes may be unable to receive the final decision, blocking lock release
•Clock Skew: Distributed timeout detection is complicated by clock differences
•Ordered Release: May need to release locks in specific order across nodes to prevent distributed deadlocks

Modern Approaches

Modern distributed databases like CockroachDB, Spanner, and YugabyteDB use sophisticated variations of 2PC with features like parallel commits, pipelining, and optimistic lock release to minimize the performance impact of distributed lock coordination. These systems maintain 2PL semantics while achieving high throughput across distributed nodes.

Shrinking Phase

The Art of Letting Go

Mechanics of Lock Release

When Does the Shrinking Phase Begin?

Lock Release Ordering

The Danger of Early Release

Implicit Release at Commit and Abort

Shrinking Phase and System Concurrency

Lock Release in Distributed Systems

There’s more to understand.

Shrinking Phase

The Art of Letting Go

Mechanics of Lock Release

When Does the Shrinking Phase Begin?

Lock Release Ordering

The Danger of Early Release

Implicit Release at Commit and Abort

Shrinking Phase and System Concurrency

Lock Release in Distributed Systems

There’s more to understand.