Strong Consistency: The Gold Standard of Distributed Correctness

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
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81

# Example: Testing for Linearizability
 
# Client A and Client B interact with a register
# Initial value: x = 0
 
# Timeline:
# ─────────────────────────────────────────────────────────────────
# Time     0    1    2    3    4    5    6    7    8
# ─────────────────────────────────────────────────────────────────
# Client A: [──── write(x=1) ────]
#                              ↑ response at time 3
#
# Client B:                [──── read(x) ────]
#                          ↑ invocation at time 2
#                                          ↑ response at time 5
# ─────────────────────────────────────────────────────────────────
 
# QUESTION: What can Client B's read return?
 
# Analysis:
# - Client A's write spans time [0, 3]
# - Client B's read spans time [2, 5]  
# - These operations OVERLAP (time 2-3)
 
# In a linearizable system:
# - If write's linearization point is before read's: read returns 1
# - If read's linearization point is before write's: read returns 0
# - BOTH are valid! (because operations overlap)
 
# What WOULD violate linearizability:
# - If write completes (response at time 3)
# - Then read STARTS (invocation at time 4)
# - Read returns 0
# This violates linearizability because there's no possible
# linearization order where read comes before write.
 
class LinearizabilityChecker:
    """
    Simplified linearizability checker for educational purposes.
    Real checkers (like Jepsen's Knossos) are far more complex.
    """
    
    def __init__(self):
        self.history = []  # List of (op_type, op_id, value, start, end)
    
    def add_operation(self, op_type: str, op_id: str, 
                      value, start_time: float, end_time: float):
        """Record an operation with its time bounds."""
        self.history.append({
            'type': op_type,
            'id': op_id,
            'value': value,
            'start': start_time,
            'end': end_time
        })
    
    def check_linearizable(self) -> bool:
        """
        Attempt to find a linearization of the history.
        Returns True if a valid linearization exists.
        
        This is NP-complete in general, so we use heuristics
        for real-world checking.
        """
        # Sort by end time (partial order)
        sorted_ops = sorted(self.history, key=lambda x: x['end'])
        
        # Try to find valid linearization points
        # (Simplified: actual algorithm is much more complex)
        for op in sorted_ops:
            # Each operation must have a linearization point
            # within its [start, end] interval
            if not self._can_linearize(op):
                return False
        return True
    
    def _can_linearize(self, op) -> bool:
        """Check if operation can be assigned a valid linearization point."""
        # Implementation would check all ordering constraints
        # This is where the NP-complete complexity lives
        return True  # Placeholder

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
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106

# Consensus-Based Strong Consistency
# Simplified view of how Raft provides linearizability
 
class RaftNode:
    """
    Simplified Raft node demonstrating strong consistency.
    Real implementations are significantly more complex.
    """
    
    def __init__(self, node_id: str, cluster_nodes: list):
        self.node_id = node_id
        self.cluster = cluster_nodes
        self.state = "follower"  # follower, candidate, leader
        self.current_term = 0
        self.log = []  # List of (term, index, command)
        self.commit_index = 0
        self.state_machine = {}  # Actual data storage
    
    def write(self, key: str, value: any) -> bool:
        """
        Strong consistency write operation.
        Only leader can accept writes.
        """
        if self.state != "leader":
            # Redirect to leader
            return self._forward_to_leader("write", key, value)
        
        # Step 1: Append to local log (not committed yet)
        entry = {
            'term': self.current_term,
            'index': len(self.log),
            'command': {'type': 'write', 'key': key, 'value': value}
        }
        self.log.append(entry)
        
        # Step 2: Replicate to followers
        # Must get acknowledgment from MAJORITY before committing
        acks = 1  # Count self
        quorum = len(self.cluster) // 2 + 1
        
        for node in self.cluster:
            if node != self.node_id:
                # Send AppendEntries RPC
                if self._send_append_entries(node, entry):
                    acks += 1
        
        # Step 3: If majority acknowledged, commit
        if acks >= quorum:
            self.commit_index = entry['index']
            # Apply to state machine
            self.state_machine[key] = value
            
            # Notify followers to commit
            self._broadcast_commit(entry['index'])
            return True
        
        # Failed to reach quorum - DO NOT acknowledge to client
        return False
    
    def read(self, key: str) -> any:
        """
        Linearizable read operation.
        Multiple strategies possible.
        """
        if self.state == "leader":
            # Option A: Read from leader
            # Must confirm we're still leader (quorum check)
            if not self._confirm_leadership():
                return self._forward_to_leader("read", key)
            
            # Safe to read - we're confirmed leader
            return self.state_machine.get(key)
        
        else:
            # Option B: Forward to leader
            return self._forward_to_leader("read", key)
            
            # Option C: Read from follower with ReadIndex protocol
            # (requires asking leader for current commit index)
    
    def _confirm_leadership(self) -> bool:
        """
        Heartbeat to quorum to confirm we're still leader.
        Prevents stale reads after network partition.
        """
        responses = 1  # Count self
        quorum = len(self.cluster) // 2 + 1
        
        for node in self.cluster:
            if node != self.node_id:
                # Send heartbeat
                if self._send_heartbeat(node):
                    responses += 1
        
        return responses >= quorum
    
    def _send_append_entries(self, node: str, entry: dict) -> bool:
        """Send log entry to follower node."""
        # Network call - may fail, timeout, etc.
        # Real implementation handles retries, term checking, etc.
        pass
    
    def _forward_to_leader(self, operation: str, *args) -> any:
        """Forward operation to current leader."""
        # Real implementation discovers leader via cluster state
        pass

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
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109

# How Google Spanner Achieves Global Strong Consistency
# Using TrueTime for External Consistency
 
class TrueTime:
    """
    Simplified model of Google's TrueTime API.
    Real TrueTime uses GPS + Atomic clocks in datacenters.
    """
    
    def now(self) -> tuple:
        """
        Returns [earliest, latest] - the interval within which
        the true current time definitely falls.
        
        Real TrueTime typically has ~7ms uncertainty (ε).
        GPS/atomic clocks bound this uncertainty.
        """
        current_time = self._get_clock_reading()
        epsilon = self._get_clock_uncertainty()  # Usually ~7ms
        
        return (current_time - epsilon, current_time + epsilon)
    
    def after(self, timestamp: float) -> bool:
        """Returns True if timestamp is definitely in the past."""
        earliest, _ = self.now()
        return earliest > timestamp
    
    def before(self, timestamp: float) -> bool:
        """Returns True if timestamp is definitely in the future."""
        _, latest = self.now()
        return latest < timestamp
 
 
class SpannerTransaction:
    """
    Simplified model of Spanner commit protocol.
    Demonstrates how TrueTime enables external consistency.
    """
    
    def __init__(self, truetime: TrueTime):
        self.truetime = truetime
        self.read_timestamp = None
        self.commit_timestamp = None
    
    def commit(self, writes: list) -> bool:
        """
        Commit transaction with external consistency guarantee.
        
        KEY INSIGHT: After commit returns, any transaction
        that starts will see this transaction's writes.
        """
        # Step 1: Acquire locks on all written keys (via Paxos groups)
        if not self._acquire_locks(writes):
            return False
        
        # Step 2: Choose commit timestamp
        # Must be >= any timestamp seen during transaction
        # Must be >= TrueTime.now().latest
        _, latest = self.truetime.now()
        self.commit_timestamp = max(
            self._max_timestamp_seen(),
            latest
        )
        
        # Step 3: THE CRITICAL WAIT
        # We must wait until we're sure the commit timestamp
        # is in the past at ALL nodes worldwide.
        #
        # Wait until: TrueTime.now().earliest > commit_timestamp
        #
        # This typically takes ~2ε (~14ms with 7ms uncertainty)
        
        while not self.truetime.after(self.commit_timestamp):
            # Wait for clock uncertainty to pass
            time.sleep(0.001)  # 1ms sleep
        
        # Step 4: Now safe to commit
        # Any future transaction will have a start timestamp
        # greater than our commit timestamp (external consistency!)
        
        self._apply_writes(writes)
        self._release_locks(writes)
        
        return True
    
    def _max_timestamp_seen(self) -> float:
        """Return maximum timestamp observed during reads."""
        # Implementation tracks timestamps of all read values
        pass
 
 
# WHY THIS WORKS:
# 
# Transaction A commits with timestamp T_a
# After commit returns, TrueTime.now().earliest > T_a (guaranteed by wait)
#
# Transaction B starts after Transaction A returns
# Transaction B's start timestamp T_b comes from TrueTime.now().latest
#
# Since B started after A committed:
#   T_b >= B's TrueTime.now().latest
#   A's TrueTime.now().earliest > T_a  (commit waited for this)
#   
# Real time ordering ensures:
#   B's TrueTime.now().latest >= A's TrueTime.now().earliest
#   Therefore: T_b > T_a
#
# Since B's timestamp > A's timestamp, and Spanner orders by timestamp,
# B will always see A's writes. External consistency achieved!

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
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159

# Implementing Linearizable Reads
# Multiple patterns with different trade-offs
 
class LinearizableReader:
    """
    Demonstrates different approaches to ensuring read linearizability.
    """
    
    def __init__(self, cluster: list, leader_lease_duration: float = 10.0):
        self.cluster = cluster
        self.lease_duration = leader_lease_duration
        self.leader_lease_start = None
        self.is_leader = False
    
    # ═══════════════════════════════════════════════════════════════
    # PATTERN 1: Read from leader with leadership confirmation
    # ═══════════════════════════════════════════════════════════════
    
    def quorum_read(self, key: str) -> any:
        """
        Confirm leadership via quorum before reading.
        Most straightforward linearizable read.
        
        Latency: ~2 RTT (1 for quorum check, 1 for actual read)
        """
        if not self.is_leader:
            return self._forward_to_leader(key)
        
        # Confirm we're still leader
        # (Could be stale after network partition)
        if not self._confirm_leadership_via_heartbeat():
            # We're not leader anymore - redirect
            return self._forward_to_leader(key)
        
        # Safe to read from local state
        return self._local_read(key)
    
    def _confirm_leadership_via_heartbeat(self) -> bool:
        """Send heartbeats to majority to confirm leadership."""
        acks = 1  # Count self
        quorum = len(self.cluster) // 2 + 1
        
        for node in self.cluster:
            if self._send_heartbeat(node):
                acks += 1
        
        return acks >= quorum
    
    # ═══════════════════════════════════════════════════════════════
    # PATTERN 2: ReadIndex Protocol (Raft-style)
    # ═══════════════════════════════════════════════════════════════
    
    def read_index_read(self, key: str) -> any:
        """
        Follower-based linearizable read using ReadIndex protocol.
        Allows distributing read load across followers.
        
        Steps:
        1. Ask leader for current committed index
        2. Wait until local log reaches that index
        3. Read from local state
        
        Latency: ~1 RTT to leader + local wait
        """
        # Step 1: Get committed index from leader
        committed_index = self._get_committed_index_from_leader()
        
        if committed_index is None:
            # Leader unreachable - fail or retry
            raise Exception("Cannot reach leader for ReadIndex")
        
        # Step 2: Wait for local log to catch up
        self._wait_for_local_commit(committed_index)
        
        # Step 3: Now safe to read locally
        return self._local_read(key)
    
    def _get_committed_index_from_leader(self) -> int:
        """Query leader for current commit index."""
        # Leader must first confirm its own leadership
        # via quorum (can piggyback on heartbeat)
        pass
    
    def _wait_for_local_commit(self, target_index: int):
        """Wait until local log is committed up to target."""
        while self._get_local_commit_index() < target_index:
            time.sleep(0.001)  # 1ms polling
    
    # ═══════════════════════════════════════════════════════════════
    # PATTERN 3: Lease-based reads
    # ═══════════════════════════════════════════════════════════════
    
    def lease_read(self, key: str) -> any:
        """
        Fast reads when leader holds valid lease.
        Requires synchronized clocks (bounded drift).
        
        Latency: ~0 RTT when lease valid, ~1-2 RTT to renew
        """
        if not self.is_leader:
            return self._forward_to_leader(key)
        
        # Check if lease is still valid
        if self._lease_is_valid():
            # Fast path: no network required
            return self._local_read(key)
        
        # Lease expired - need to renew
        if self._renew_lease():
            return self._local_read(key)
        
        # Failed to renew - no longer leader
        return self._forward_to_leader(key)
    
    def _lease_is_valid(self) -> bool:
        """
        Check if current lease is valid.
        
        CRITICAL: Lease must expire BEFORE followers might
        elect a new leader. Must account for clock drift!
        """
        if self.leader_lease_start is None:
            return False
        
        current_time = time.time()
        elapsed = current_time - self.leader_lease_start
        
        # Use conservative estimate (account for clock drift)
        max_clock_drift = 0.1  # 100ms assumed max drift
        safe_lease_duration = self.lease_duration - max_clock_drift
        
        return elapsed < safe_lease_duration
    
    # ═══════════════════════════════════════════════════════════════
    # PATTERN 4: Strict quorum reads (Dynamo-style)
    # ═══════════════════════════════════════════════════════════════
    
    def quorum_read_dynamo_style(self, key: str) -> any:
        """
        Read from majority of replicas, return highest version.
        
        For linearizability with N=3: R=2, W=2 (R + W > N)
        
        Latency: parallelized, but wait for majority
        """
        n_replicas = 3
        read_quorum = 2
        
        responses = []
        for replica in self._get_replicas(key):
            response = self._read_from_replica(replica, key)
            if response:
                responses.append(response)
        
        if len(responses) < read_quorum:
            raise Exception("Failed to reach read quorum")
        
        # Return value with highest version
        return max(responses, key=lambda r: r['version'])['value']