✍️ 필사 모드: The Complete Guide to Probabilistic Data Structures 2025: Bloom Filter, HyperLogLog, Count-Min Sketch, and Real-World Usage

TL;DR

• Essence of probabilistic data structures: Trade 100% accuracy for extreme savings in memory/speed
• Bloom Filter: "This element may be in the set" — false positive O, false negative X. Core of RocksDB, Cassandra, BigTable
• HyperLogLog: Count unique elements among billions with 12KB. Used by Redis PFADD/PFCOUNT, BigQuery, Druid
• Count-Min Sketch: Estimate element frequency in streams. Hot-key detection, DDoS detection
• Tradeoff: accuracy vs memory vs speed — explosive efficiency when exact answers are unnecessary

1. What Are Probabilistic Data Structures?

1.1 Core Idea

"If you don't need a 100% accurate answer, you can save memory and speed extremely."

Traditional data structures:

• Storing 1M elements in a HashMap → about 50 MB
• "Does this element exist?" → exact answer

Probabilistic data structures:

• Storing 1M elements in a Bloom Filter → about 1 MB (50x savings!)
• "Does this element exist?" → "maybe" or "definitely not"

Tradeoff: 50x memory savings + a small false positive rate (1%).

1.2 Why Do We Need Them?

In large-scale systems:

• Memory is expensive — RAM is 100x+ more expensive than disk
• Must fit in L1/L2 cache to be fast → smaller is better
• Billions of elements — HashMap becomes infeasible
• Distributed systems — smaller data is more network-efficient

1.3 Cases Where Accuracy Is Unnecessary

2. Bloom Filter — The Most Famous Probabilistic Data Structure

2.1 Principle

Bit array + multiple hash functions.

Insertion:

1. Pass the element through K hash functions → K indices
2. Set the bit at each index to 1

Query:

1. Same K hash functions
2. If all bits are 1 → "maybe present"
3. If any bit is 0 → "definitely not present"

2.2 Visualization

Bit array (8 bits):  [0, 0, 0, 0, 0, 0, 0, 0]

Insert "apple" (hash → [1, 4]):
                  [0, 1, 0, 0, 1, 0, 0, 0]

Insert "banana" (hash → [3, 6]):
                  [0, 1, 0, 1, 1, 0, 1, 0]

Query "apple" (hash → [1, 4]): both 1 → "maybe present" OK
Query "grape" (hash → [2, 5]): has a 0 → "definitely not present" NO
Query "cherry" (hash → [1, 6]): coincidentally both 1 → "maybe present" (false positive)

2.3 Computing the False Positive Rate

Formula:

P(fp) = (1 - e^(-k*n/m))^k

• n: number of elements
• m: number of bits
• k: number of hash functions

Optimal k:

k = (m/n) * ln(2)

If you target 1% false positive:

m/n ≈ 9.6 bits
k ≈ 7

1M elements = 1.2 MB + 7 hash functions = 50x smaller than HashMap.

2.4 Python Implementation

import hashlib
import math

class BloomFilter:
    def __init__(self, n, fp_rate):
        self.size = -int(n * math.log(fp_rate) / (math.log(2) ** 2))
        self.hash_count = int(self.size / n * math.log(2))
        self.bits = [0] * self.size

    def _hashes(self, item):
        for i in range(self.hash_count):
            h = hashlib.md5(f"{item}{i}".encode()).hexdigest()
            yield int(h, 16) % self.size

    def add(self, item):
        for h in self._hashes(item):
            self.bits[h] = 1

    def contains(self, item):
        return all(self.bits[h] for h in self._hashes(item))

# Usage
bf = BloomFilter(n=1000000, fp_rate=0.01)
bf.add("alice@example.com")
print(bf.contains("alice@example.com"))  # True
print(bf.contains("bob@example.com"))    # Almost always False

2.5 Limitations of Bloom Filters

• No deletion — setting bits to 0 would affect other elements
• Hard to resize — must create a new filter and reinsert all elements
• No count — cannot know how many elements are inside

Solutions: Counting Bloom Filter, Cuckoo Filter (see below).

2.6 Real-World Usage of Bloom Filters

RocksDB / Cassandra / BigTable

In LSM-tree-based DBs, there is a Bloom Filter per SSTable:

Query: "key123"

SSTable 1 Bloom Filter: "definitely not present" → skip disk read OK
SSTable 2 Bloom Filter: "definitely not present" → skip disk read OK
SSTable 3 Bloom Filter: "maybe present"         → read disk (actual search)

Effect: 99% reduction in disk I/O. Entire SSTables are never read.

CDN Cache

Request: GET /image/123.jpg

Bloom Filter: "not in cache" → go straight to origin (skip cache lookup)
              "maybe"        → cache lookup

Effect: Saves time on cache miss.

Password Leak Checks

HaveIBeenPwned's 100M+ leaked password database → clients can check locally (download the Bloom Filter).

Chrome Malicious URL Check

Google's known malicious URL database → sent to clients as a Bloom Filter → browser checks immediately.

3. Counting Bloom Filter

3.1 Solving the Problem

A basic Bloom Filter does not support deletion. A Counting Bloom Filter uses counters instead of bits:

Array: [0, 0, 0, 0, 0, 0, 0, 0]  (each cell is a 4-bit counter)

Insert "apple" ([1, 4]):
        [0, 1, 0, 0, 1, 0, 0, 0]

Insert "apple" again:
        [0, 2, 0, 0, 2, 0, 0, 0]

Delete "apple":
        [0, 1, 0, 0, 1, 0, 0, 0]

3.2 Drawbacks

• 4x the memory — 1 bit → 4 bits (to prevent overflow)
• Still has false positives

3.3 Cuckoo Filter (Alternative)

The Cuckoo Filter is a superior alternative to the Bloom Filter:

• Supports deletion
• Lower false positive rate
• Comparable memory efficiency
• Based on "Cuckoo hashing"

# pycuckoofilter example
from cuckoofilter import CuckooFilter

cf = CuckooFilter(capacity=1000000, error_rate=0.01)
cf.insert("alice")
print(cf.contains("alice"))  # True
cf.delete("alice")  # Possible!

When Cuckoo > Bloom: when deletion is needed, or when a lower false positive rate is required.

4. HyperLogLog — Cardinality Estimation

4.1 The Problem

"How many unique visitors came to our site today?"

Traditional method: Store all visitor IDs in a HashSet → 10M visitors = 800 MB.

HyperLogLog: Estimate up to billions of elements with 12 KB. 1% error.

4.2 Basic Idea

Hash each element → track the number of leading zeros in the binary representation.

"alice" → hash → 00000010110...  (5 leading zeros)
"bob"   → hash → 00100110...     (2 leading zeros)
"carol" → hash → 0001110...      (3 leading zeros)

Observation: The probability of K leading zeros in a random hash is 1/2^K.

→ "We saw at most 5 leading zeros" → estimated about 2^5 = 32 unique elements.

Reality: Single measurements are inaccurate. Split across multiple buckets and average → improved accuracy. This is HyperLogLog.

4.3 Accuracy

Key point: Memory is independent of the number of elements! 100M elements and 1M elements take the same memory.

4.4 Redis HyperLogLog

PFADD visitors:2025-04-15 "user1" "user2" "user3"
PFADD visitors:2025-04-15 "user1"  # duplicates ignored
PFCOUNT visitors:2025-04-15  # about 3 (could be 2 or 3)

# Merge multiple HLLs
PFMERGE visitors:week visitors:2025-04-14 visitors:2025-04-15

4.5 Real-World Usage of HyperLogLog

Daily Active Users (DAU)

import redis

r = redis.Redis()

def track_user(user_id):
    today = datetime.now().strftime("%Y-%m-%d")
    r.pfadd(f"dau:{today}", user_id)

def get_dau(date):
    return r.pfcount(f"dau:{date}")

Hundreds of millions of users can be handled with 12 KB.

Unique IP Count (DDoS Analysis)

PFADD unique_ips:2025-04-15 "1.2.3.4" "5.6.7.8"
PFCOUNT unique_ips:2025-04-15

Big Data — BigQuery, Druid

SELECT APPROX_COUNT_DISTINCT(user_id) FROM events;

Uses HyperLogLog internally. 100x+ faster than exact COUNT(DISTINCT).

5. Count-Min Sketch — Frequency Estimation

5.1 The Problem

In a stream, "how many times has each element appeared?"

Traditional: HashMap → memory explosion.

Count-Min Sketch: Estimate frequencies with small memory.

5.2 Principle

2D array (d × w) + d hash functions.

Insert:

• Compute an index with each hash function
• Increment the corresponding counter

Query:

• Same hash functions
• Return the minimum counter value among them (hence "Count-Min")

5.3 Visualization

Array (3 x 5):
[0, 0, 0, 0, 0]  ← hash1
[0, 0, 0, 0, 0]  ← hash2
[0, 0, 0, 0, 0]  ← hash3

Insert "apple" 5 times:
hash1("apple") = 1 → counter [0, 5, 0, 0, 0]
hash2("apple") = 3 → counter [0, 0, 0, 5, 0]
hash3("apple") = 0 → counter [5, 0, 0, 0, 0]

Insert "banana" 2 times:
hash1("banana") = 4 → [0, 5, 0, 0, 2]
hash2("banana") = 1 → [0, 2, 0, 5, 0]
hash3("banana") = 2 → [5, 0, 2, 0, 0]

Query "apple": min(5, 5, 5) = 5 OK
Query "banana": min(2, 2, 2) = 2 OK

On collision: if another element collides, the counter may be inflated, but since we take the minimum, it is always >= the true count (overestimate).

5.4 Accuracy

Error ε, confidence 1-δ
Required size:
  w = ⌈e/ε⌉
  d = ⌈ln(1/δ)⌉

Example: 1% error, 99% confidence → w=272, d=5 → 1360 counters = 5 KB. Can handle billions of elements.

5.5 Real-World Usage

Heavy Hitters (Hot Key Detection)

# Find the most frequent keys
sketch = CountMinSketch(width=272, depth=5)

for query in stream:
    sketch.add(query)

# Top-K discovery (uses a separate min-heap)

Alternative to Memcached LRU

The tinylfu cache uses Count-Min Sketch to track frequency → keeps only frequently used keys in cache.

DDoS Detection

Estimate the request count per IP → block when exceeding a threshold.

Apache Spark, Flink

Frequency analysis in large-scale stream processing.

6. MinHash — Jaccard Similarity

6.1 The Problem

"How similar are two sets?" (Jaccard similarity)

J(A, B) = |A ∩ B| / |A ∪ B|

Traditional method: compare all elements → O(|A| × |B|).

6.2 MinHash Principle

The probability that the "minimum hash value" of each set is equal = Jaccard similarity.

def minhash(set_, num_hashes=128):
    return [min(hash(item, seed=i) for item in set_) for i in range(num_hashes)]

# Compare MinHashes of two sets
mh1 = minhash(set1)
mh2 = minhash(set2)

similarity = sum(a == b for a, b in zip(mh1, mh2)) / len(mh1)
# similarity ≈ Jaccard(set1, set2)

6.3 Usage

Duplicate Document Detection

Finding near-identical pages among billions of web pages. Used by Google.

Plagiarism Detection

Comparison of papers and code.

Recommendation Systems

Finding "similar users".

Clustering

Combined with LSH (Locality Sensitive Hashing) to group similar items.

7. Skip List — Probabilistic Balanced Tree

7.1 Sorted Data Structure

A probabilistic alternative to balanced binary trees (AVL, Red-Black).

7.2 Principle

Multi-level linked list. Upper levels are sparse, lower levels dense.

Level 3:  1 ───────────────→ 9
Level 2:  1 ────→ 4 ────→ 7 ─→ 9
Level 1:  1 → 3 → 4 → 5 → 7 → 8 → 9

Search: start from the top, drop a level if too large.

Level on insert: decided by coin flip. 50% chance of going to the next level.

7.3 Performance

Advantages: simpler to implement than balanced trees, easy lock-free concurrency.

7.4 Usage

• Redis Sorted Set: internal structure of ZADD/ZRANGE
• LevelDB/RocksDB memtable
• Java ConcurrentSkipListMap

8. T-Digest — Quantile Estimation

8.1 The Problem

How do you compute quantiles like p99 latency in a large-scale stream?

Traditional: store all values then sort → memory explosion.

T-Digest: maintain a "sketch" of the distribution with small memory.

8.2 Principle

Group values into buckets. More buckets at the distribution's tail (p99, p99.9), fewer in the middle (p50).

Rationale: we usually care about the tail (tail latency). The median can be approximate.

8.3 Usage

from tdigest import TDigest

t = TDigest()
for value in stream:
    t.update(value)

p50 = t.percentile(50)
p95 = t.percentile(95)
p99 = t.percentile(99)
p999 = t.percentile(99.9)

8.4 Usage

• Monitoring: Prometheus, Datadog, New Relic
• Databases: Druid, Apache Pinot
• Stream processing: Apache Beam, Flink

9. Tradeoff Comparison

9.1 Comparison of All Data Structures

9.2 Memory Comparison (100M Unique Users)

HyperLogLog is 300,000x smaller.

10. Real-World — Heavy Hitter Detection in Kafka

10.1 Scenario

1M messages/sec on a Kafka topic. "Which user_id is most active?" (spam detection)

10.2 Implementation

from countminsketch import CountMinSketch
import heapq

# Count-Min Sketch + Top-K heap
sketch = CountMinSketch(width=10000, depth=5)
top_k = []  # min-heap of (count, user_id)
TOP_K_SIZE = 100

def process_message(user_id):
    sketch.add(user_id)
    estimated_count = sketch.estimate(user_id)
    
    if len(top_k) < TOP_K_SIZE:
        heapq.heappush(top_k, (estimated_count, user_id))
    elif estimated_count > top_k[0][0]:
        heapq.heapreplace(top_k, (estimated_count, user_id))

# Memory: ~200 KB handles 100M users

10.3 Result

Traditional HashMap = several GB → CMS = 200 KB (10,000x savings).

Accepting a bit of inaccuracy yields explosive memory efficiency.

Quiz

References

• Bloom Filter Original Paper — Burton Bloom (1970)
• HyperLogLog Original Paper — Flajolet et al.
• Count-Min Sketch Original Paper — Cormode & Muthukrishnan
• Cuckoo Filter — Fan et al.
• Probabilistic Data Structures and Algorithms — Andrii Gakhov (book)
• Redis HyperLogLog
• RocksDB Bloom Filter
• Apache DataSketches — high-quality implementations
• stream-lib — Java library
• t-digest — Ted Dunning's implementation
• pdsa.gakhov.com — visualizations + simulations