Cassandra Data Modeling Anti-Patterns

These patterns work fine in development and testing. They might even work in early production. Then data grows, traffic increases, and they fail—usually in ways that are not obvious without understanding Cassandra internals.

Unbounded partitions grow until reads timeout. Tombstone accumulation makes queries slower over time. ALLOW FILTERING queries that took milliseconds start taking minutes. Queue patterns create tombstones faster than compaction can remove them.

The fix is usually straightforward once the problem is understood. This guide covers each anti-pattern: why it seems reasonable, how it fails, and how to fix it.

Anti-Pattern Severity Reference

Anti-Pattern	Severity	Time to Impact	Symptoms
Unbounded partitions	Critical	Weeks-Months	Timeouts, OOM, repair failures
Tombstone accumulation	Critical	Days-Weeks	Increasing read latency, query failures
ALLOW FILTERING	Critical	Immediate	Full cluster scan, timeouts
Queue pattern	High	Days	Tombstones, hot partitions
Secondary index abuse	High	Immediate	All-node queries, high latency
Large IN clauses	High	Immediate	Coordinator overload
Collection abuse	Medium	Weeks	Memory pressure, slow writes
Counter misuse	Medium	Immediate	Errors, data inconsistency
Over-normalization	Medium	Immediate	Multiple queries per operation

---

Unbounded Partitions (Critical)

The Problem

Partitions that grow without limit eventually exceed what Cassandra can handle efficiently.

-- ANTI-PATTERN: Partition grows forever
CREATE TABLE user_activity (
    user_id UUID,
    activity_time TIMESTAMP,
    activity_type TEXT,
    details TEXT,
    PRIMARY KEY ((user_id), activity_time)
);

What happens over time:

Month 1:   1,000 rows per user        ~100 KB per partition ✓
Month 6:   50,000 rows per user       ~5 MB per partition ✓
Year 1:    500,000 rows per user      ~50 MB per partition ⚠
Year 3:    5,000,000 rows per user    ~500 MB per partition ✗
Year 5:    15,000,000 rows per user   ~1.5 GB per partition ☠

What Goes Wrong

Stage	Partition Size	Symptoms
1. Read Performance Degrades	> 100 MB	Queries slow down. "Get latest activity" now scans gigabytes to find recent rows. Bloom filters help but not enough.
2. Compaction Problems	> 500 MB	Compaction must load entire partition into memory. Compaction threads consume heap, leading to GC pauses. Other operations stall waiting for compaction to complete.
3. Repair Failures	> 1 GB	Repair streams entire partitions between nodes. Streaming 1GB over network causes timeouts. Repair fails repeatedly. Data consistency cannot be maintained.
4. Cluster Instability	> 2 GB	Nodes hosting these partitions become unresponsive. Hints queue up for unavailable nodes. Cascade failures begin. Full cluster becomes unstable.

Detection

# Check partition size distribution
nodetool tablehistograms keyspace.user_activity

# Warning signs in output:
# 99th percentile partition size > 100MB
# Max partition size > 500MB

# Check for large partition warnings in logs
grep "Compacting large partition" /var/log/cassandra/system.log
grep "Writing large partition" /var/log/cassandra/system.log

Solution: Time Bucketing

-- CORRECT: Bounded partitions by time bucket
CREATE TABLE user_activity (
    user_id UUID,
    month TEXT,              -- '2024-01'
    activity_time TIMESTAMP,
    activity_type TEXT,
    details TEXT,
    PRIMARY KEY ((user_id, month), activity_time)
) WITH CLUSTERING ORDER BY (activity_time DESC);

Now each partition covers one month:

user_id=alice, month='2024-01': 40,000 rows, ~4 MB ✓
user_id=alice, month='2024-02': 38,000 rows, ~3.8 MB ✓
user_id=alice, month='2024-03': 42,000 rows, ~4.2 MB ✓

Migration Strategy

For existing unbounded partitions:

-- 1. Create new time-bucketed table
CREATE TABLE user_activity_v2 (...);

-- 2. Migrate data in application
-- Read from old table, write to new with bucket

-- 3. Dual-write during transition
-- Write to both tables

-- 4. Switch reads to new table

-- 5. Drop old table after validation

---

Tombstone Accumulation (Critical)

The Problem

Every DELETE in Cassandra creates a tombstone—a marker that says "this data was deleted." Tombstones must be scanned during reads until they're purged by compaction.

-- ANTI-PATTERN: High-frequency deletes
CREATE TABLE message_queue (
    queue_id TEXT,
    message_id TIMEUUID,
    payload TEXT,
    PRIMARY KEY ((queue_id), message_id)
);

-- Application pattern
INSERT INTO message_queue (...) VALUES (...);
-- Process message
DELETE FROM message_queue WHERE queue_id = ? AND message_id = ?;
-- Repeat thousands of times per hour

What Goes Wrong

Day 1:   100 live rows, 1,000 tombstones
         Read scans 1,100 entries to return 100 rows
         Query time: 5ms

Day 7:   100 live rows, 50,000 tombstones
         Read scans 50,100 entries to return 100 rows
         Query time: 200ms

Day 30:  100 live rows, 500,000 tombstones
         Read scans 500,100 entries to return 100 rows
         Query time: 2,000ms → TIMEOUT

         Eventually: TombstoneOverwhelmingException
         Query fails at tombstone_failure_threshold (default 100,000)

Why Tombstones Persist

Tombstones cannot be removed until:

1. gc_grace_seconds has passed (default 10 days)
2. All replicas have the tombstone (ensured by repair)
3. Compaction runs and merges SSTables containing the tombstone

Day	Event	Notes
0	DELETE executed	Tombstone created
1-9	Tombstone must stay	gc_grace_seconds = 10 days
10	Tombstone eligible for removal	Requires compaction to actually remove
??	Compaction runs	Tombstone finally removed

Tombstone Removal Risks

• If repair has not run: Tombstone removal is unsafe (data resurrection risk)
• If compaction is behind: Tombstone persists indefinitely

Detection

-- Enable query tracing
TRACING ON;
SELECT * FROM message_queue WHERE queue_id = 'orders';

-- Look for in trace output:
-- "Read 100 live rows and 50000 tombstone cells"

# Check tombstone metrics
nodetool tablestats keyspace.message_queue | grep -i tombstone

# Check for warnings in logs
grep "tombstone" /var/log/cassandra/system.log
grep "TombstoneOverwhelmingException" /var/log/cassandra/system.log

# cassandra.yaml thresholds
tombstone_warn_threshold: 1000     # Log warning
tombstone_failure_threshold: 100000  # Fail query

Solutions

Solution 1: Use TTL instead of DELETE

-- Let data expire automatically
INSERT INTO message_queue (queue_id, message_id, payload)
VALUES (?, now(), ?)
USING TTL 3600;  -- Expires after 1 hour

-- No explicit DELETE needed
-- TTL creates tombstones too, but predictably

Solution 2: Time-windowed tables with TWCS

CREATE TABLE message_queue (
    queue_id TEXT,
    hour TIMESTAMP,
    message_id TIMEUUID,
    payload TEXT,
    PRIMARY KEY ((queue_id, hour), message_id)
) WITH compaction = {
    'class': 'TimeWindowCompactionStrategy',
    'compaction_window_size': '1',
    'compaction_window_unit': 'HOURS'
}
AND default_time_to_live = 86400
AND gc_grace_seconds = 3600;

-- TWCS drops entire SSTables when all data expires
-- No tombstone accumulation

Solution 3: Reduce gc_grace_seconds (with frequent repair)

-- Only if repair runs more frequently than gc_grace_seconds
ALTER TABLE message_queue WITH gc_grace_seconds = 86400;  -- 1 day

-- MUST run repair daily or more frequently
-- Otherwise: data resurrection risk

Solution 4: Range deletes for batch cleanup

-- Instead of individual deletes
-- Accumulate work, then delete entire partition

-- Delete all messages older than 1 hour (partition delete)
DELETE FROM message_queue
WHERE queue_id = 'orders'
  AND message_id < minTimeuuid('2024-01-15 14:00:00');

---

ALLOW FILTERING (Critical)

The Problem

ALLOW FILTERING tells Cassandra to scan all partitions in a table, filtering results in memory.

-- ANTI-PATTERN: Query without partition key
SELECT * FROM users WHERE country = 'US' ALLOW FILTERING;

What Actually Happens

For SELECT * FROM users WHERE country = 'US' ALLOW FILTERING:

Step	Action
1	Coordinator receives query
2	Coordinator sends request to all nodes: "Scan ALL your partitions, filter by country"
3	Each node reads every SSTable, deserializes every row, checks if country = 'US'
4	Each node returns matching rows to coordinator
5	Coordinator aggregates and returns results

Complexity: O(n) where n = total rows in table

Table Size	Rows Scanned
1 million users	1 million rows
10 million users	10 million rows
100 million users	100 million rows

When ALLOW FILTERING Is Acceptable

In very limited cases:

-- OK: Small table (< 10,000 rows), infrequent queries
SELECT * FROM configuration WHERE feature = 'dark_mode' ALLOW FILTERING;

-- OK: Already constrained to single partition
SELECT * FROM user_orders
WHERE user_id = '550e8400-e89b-12d3-a456-426614174000'
  AND status = 'pending'
ALLOW FILTERING;
-- Only scans one partition, not entire table

Solutions

Solution 1: Create a table for the query

-- Instead of: SELECT * FROM users WHERE country = 'US' ALLOW FILTERING

-- Create dedicated table
CREATE TABLE users_by_country (
    country TEXT,
    user_id UUID,
    username TEXT,
    email TEXT,
    PRIMARY KEY ((country), user_id)
);

-- Query without ALLOW FILTERING
SELECT * FROM users_by_country WHERE country = 'US' LIMIT 100;

Solution 2: Secondary index (low-cardinality only)

-- For low-cardinality columns (< 100 unique values)
CREATE INDEX ON users (status);  -- active/inactive/pending

-- Query uses index
SELECT * FROM users WHERE status = 'active';

-- WARNING: Still queries all nodes, but more efficiently

Solution 3: SAI (Cassandra 5.0+)

-- Storage-Attached Index for high-cardinality
CREATE CUSTOM INDEX ON users (email)
USING 'StorageAttachedIndex';

SELECT * FROM users WHERE email = 'alice@example.com';

---

Queue Anti-Pattern (High)

The Problem

Using Cassandra as a message queue combines multiple anti-patterns: tombstone accumulation, hot partitions, and race conditions.

-- ANTI-PATTERN: Queue table
CREATE TABLE job_queue (
    queue_name TEXT,
    job_id TIMEUUID,
    payload TEXT,
    status TEXT,
    PRIMARY KEY ((queue_name), job_id)
);

-- Worker pattern
WHILE true:
    -- Poll for jobs
    SELECT * FROM job_queue
    WHERE queue_name = 'processing'
    LIMIT 10;

    -- Process job

    -- Delete completed job
    DELETE FROM job_queue
    WHERE queue_name = 'processing' AND job_id = ?;

What Goes Wrong

Problem	Description
1. Tombstone Accumulation	Every DELETE creates a tombstone. After 10,000 jobs: 10 live rows, 10,000 tombstones. SELECT scans 10,010 entries to return 10 rows.
2. Hot Partition	All workers read from same partition (queue_name = 'processing'). One partition handles 100% of queue traffic. Node hosting this partition is overwhelmed while others idle.
3. Race Conditions	Multiple workers poll simultaneously. Worker A and Worker B both read job_id = 123. Both process the same job → data corruption or duplicate processing.
4. No Ordering Guarantees	Cassandra's eventual consistency means DELETE might not be visible immediately, job might be re-read before delete propagates, and FIFO ordering is not guaranteed.

Solution: Use a Real Message Queue

Cassandra is not a message queue. Use:

• Apache Kafka - High-throughput streaming
• RabbitMQ - Traditional message broker
• Amazon SQS - Managed queue service
• Redis Streams - Fast in-memory queue

If Cassandra Must Be Used for Queue-Like Patterns

-- Time-windowed processing batches
CREATE TABLE processing_batches (
    batch_hour TIMESTAMP,
    worker_id INT,           -- Partition by worker to avoid contention
    job_id TIMEUUID,
    payload TEXT,
    PRIMARY KEY ((batch_hour, worker_id), job_id)
) WITH default_time_to_live = 86400;  -- TTL instead of delete

-- Each worker owns specific partition
-- Worker 0 reads (batch_hour, worker_id=0)
-- Worker 1 reads (batch_hour, worker_id=1)
-- No contention, no deletes, bounded partitions

-- Status tracking without deletes
CREATE TABLE job_status (
    job_id UUID PRIMARY KEY,
    status TEXT,
    updated_at TIMESTAMP
) WITH default_time_to_live = 604800;  -- 7 days

-- Update status instead of delete
UPDATE job_status SET status = 'completed', updated_at = toTimestamp(now())
WHERE job_id = ?;

---

Secondary Index Abuse (High)

The Problem

Secondary indexes in Cassandra query all nodes in the cluster, making them expensive for primary access patterns.

-- ANTI-PATTERN: Index as primary access pattern
CREATE TABLE users (
    user_id UUID PRIMARY KEY,
    email TEXT,
    username TEXT,
    country TEXT
);

CREATE INDEX ON users (email);

-- 90% of queries use this index
SELECT * FROM users WHERE email = ?;

What Happens

Query Type	Nodes Contacted	Latency
PRIMARY KEY query	1-3 nodes (based on RF)	2-5 ms
Secondary index query	ALL nodes	20-200 ms (scales with cluster size)

Cluster Size	Nodes Queried (Secondary Index)
5 nodes	5 nodes
50 nodes	50 nodes
500 nodes	500 nodes

When Secondary Indexes Are OK

-- OK: Low cardinality, occasional queries
CREATE INDEX ON orders (status);  -- Few values: pending, shipped, delivered

SELECT * FROM orders WHERE status = 'pending';
-- Acceptable for dashboard/reporting queries

-- OK: Combined with partition key
SELECT * FROM user_orders
WHERE user_id = ? AND status = 'pending';
-- Index only scans within one partition

When to Avoid Secondary Indexes

-- BAD: High cardinality
CREATE INDEX ON users (email);      -- Millions of unique values
CREATE INDEX ON orders (order_id);  -- Every order has unique ID

-- BAD: Primary access pattern
-- If > 10% of queries use the index, create a dedicated table

-- BAD: Large result sets
SELECT * FROM users WHERE country = 'US';  -- Returns millions of rows

Solution: Dedicated Table

-- Instead of index on email
CREATE TABLE users_by_email (
    email TEXT PRIMARY KEY,
    user_id UUID,
    username TEXT
);

-- Maintain both tables
BEGIN BATCH
    INSERT INTO users (user_id, email, username) VALUES (?, ?, ?);
    INSERT INTO users_by_email (email, user_id, username) VALUES (?, ?, ?);
APPLY BATCH;

-- Query by email hits single partition
SELECT user_id FROM users_by_email WHERE email = ?;

---

Large IN Clauses (High)

The Problem

Large IN clauses create parallel requests that overwhelm coordinators.

-- ANTI-PATTERN: Hundreds of values
SELECT * FROM products WHERE product_id IN (
    '550e8400-e89b-12d3-a456-426614174000',
    '550e8400-e89b-12d3-a456-426614174001',
    -- ... 500 more UUIDs ...
);

What Happens

For an IN clause with 500 values:

Step	Action
1	Coordinator receives query
2	Spawns 500 sub-queries (one per value)
3	Each sub-query routes to different partition on different node(s)
4	Coordinator tracks 500 outstanding requests in memory
5	All 500 must complete before response sent

Problem	Impact
Coordinator heap pressure	Memory: 500 × (request state + response buffer)
Single slow sub-query	Entire query times out
Any failure	No partial results returned
Latency	= max(all sub-query latencies)

Solutions

Solution 1: Batch in application

def get_products_batch(session, product_ids, batch_size=20):
    """Query in smaller batches."""
    results = []

    for i in range(0, len(product_ids), batch_size):
        batch = product_ids[i:i + batch_size]
        rows = session.execute(
            "SELECT * FROM products WHERE product_id IN %s",
            (tuple(batch),)
        )
        results.extend(rows)

    return results

Solution 2: Parallel async queries

async def get_products_async(session, product_ids):
    """Individual queries, controlled parallelism."""
    stmt = session.prepare("SELECT * FROM products WHERE product_id = ?")

    # Control concurrency with semaphore
    semaphore = asyncio.Semaphore(50)

    async def fetch_one(pid):
        async with semaphore:
            return await session.execute_async(stmt, [pid])

    tasks = [fetch_one(pid) for pid in product_ids]
    results = await asyncio.gather(*tasks)
    return [row for result in results for row in result]

Solution 3: Denormalize if IDs come from another query

-- If the pattern is: Get orders → Get products for those orders
-- Consider storing product info in orders table

CREATE TABLE orders_with_products (
    order_id UUID,
    product_id UUID,
    product_name TEXT,      -- Denormalized
    product_price DECIMAL,  -- Denormalized
    quantity INT,
    PRIMARY KEY ((order_id), product_id)
);

-- Single query returns order with product details

Guidelines

IN Clause Size	Recommendation
1-20	Generally safe
20-100	Monitor coordinator latency
100-500	Split into batches
500+	Redesign approach

---

Collection Abuse (Medium)

The Problem

Collections (LIST, SET, MAP) are stored as single cells and read/written atomically.

-- ANTI-PATTERN: Unbounded collection
CREATE TABLE users (
    user_id UUID PRIMARY KEY,
    name TEXT,
    activity_log LIST<TEXT>  -- Grows forever
);

-- Every page view appends to list
UPDATE users SET activity_log = activity_log + ['viewed product X']
WHERE user_id = ?;

What Goes Wrong

Month 1:    100 activities     → Collection size: 10 KB
Month 6:    50,000 activities  → Collection size: 5 MB
Year 1:     500,000 activities → Collection size: 50 MB

Problems:
- Entire 50 MB collection read on any access
- Entire 50 MB collection written on any update
- Cannot query/page through collection
- Cannot TTL individual elements
- Memory pressure during serialization

Collection Limits

Limit	Issue
~64KB	Cell size warning in logs
~10MB	Significant performance impact
~100 elements	Recommended maximum for good performance

Solution: Use a Table Instead

-- CORRECT: Separate table for unbounded data
CREATE TABLE user_activity (
    user_id UUID,
    activity_time TIMESTAMP,
    activity TEXT,
    PRIMARY KEY ((user_id), activity_time)
) WITH CLUSTERING ORDER BY (activity_time DESC)
  AND default_time_to_live = 2592000;  -- 30 days

-- Can query specific ranges
SELECT * FROM user_activity
WHERE user_id = ?
LIMIT 100;

-- Can TTL individual rows
-- Can page through results
-- Each row is independent

When Collections Are OK

-- OK: Small, bounded collections
CREATE TABLE products (
    product_id UUID PRIMARY KEY,
    name TEXT,
    tags SET<TEXT>,              -- 5-20 tags max
    attributes MAP<TEXT, TEXT>,   -- 10-20 attributes max
    image_urls LIST<TEXT>         -- 5-10 images max
);

-- Application enforces size limits
-- Collection rarely updated after initial write

---

Counter Misuse (Medium)

The Problem

Counters have strict limitations that cause errors when violated.

-- ANTI-PATTERN: Mixing counters with non-counters
CREATE TABLE page_stats (
    page_id UUID PRIMARY KEY,
    title TEXT,           -- Non-counter
    view_count COUNTER    -- Counter
);
-- ERROR: Cannot mix counter and non-counter columns

-- ANTI-PATTERN: TTL on counters
UPDATE page_views USING TTL 3600
SET views = views + 1 WHERE page_id = ?;
-- ERROR: TTL is not supported on counter mutations

-- ANTI-PATTERN: Setting counter to specific value
UPDATE page_views SET views = 0 WHERE page_id = ?;
-- ERROR: Cannot set counter to specific value

Counter Rules

Rule	Restriction
Table contents	Counter tables can ONLY contain primary key columns and counter columns. No other column types allowed.
Operations	Counters can ONLY be incremented (views = views + 1) or decremented (views = views - 1). Cannot be set to a specific value.
Limitations	No TTL allowed. No lightweight transactions. No batches with non-counter operations.
Consistency	Eventual consistency (may show stale values). Not suitable for exact counts. Good for approximate metrics.

Correct Counter Usage

-- Counter table (counters only)
CREATE TABLE page_views (
    page_id UUID PRIMARY KEY,
    view_count COUNTER,
    unique_visitors COUNTER
);

-- Metadata table (separate)
CREATE TABLE page_info (
    page_id UUID PRIMARY KEY,
    title TEXT,
    url TEXT,
    created_at TIMESTAMP
);

-- Update counter
UPDATE page_views SET view_count = view_count + 1
WHERE page_id = ?;

-- Batch counter updates (counter-only batch)
BEGIN COUNTER BATCH
    UPDATE page_views SET view_count = view_count + 1 WHERE page_id = ?;
    UPDATE page_views SET unique_visitors = unique_visitors + 1 WHERE page_id = ?;
APPLY BATCH;

When to Avoid Counters

If the application needs:

• Exact counts → Use application-level aggregation
• Time-windowed counts → Use pre-aggregated tables
• TTL on counts → Use regular columns with application increment
• Resettable counts → Use regular BIGINT columns

-- Alternative: Manual aggregation
CREATE TABLE daily_page_views (
    page_id UUID,
    date DATE,
    views BIGINT,
    PRIMARY KEY ((page_id), date)
);

-- Can set specific values, use TTL
UPDATE daily_page_views
SET views = views + 1  -- Application handles read-modify-write
WHERE page_id = ? AND date = toDate(now());

---

Over-Normalization (Medium)

The Problem

Applying relational normalization to Cassandra requires multiple queries and application-level joins.

-- ANTI-PATTERN: Normalized schema
CREATE TABLE users (
    user_id UUID PRIMARY KEY,
    username TEXT,
    email TEXT
);

CREATE TABLE addresses (
    address_id UUID PRIMARY KEY,
    user_id UUID,
    street TEXT,
    city TEXT,
    country TEXT
);

CREATE TABLE orders (
    order_id UUID PRIMARY KEY,
    user_id UUID,
    total DECIMAL,
    created_at TIMESTAMP
);

-- To display order with user and address:
-- Query 1: Get order
-- Query 2: Get user
-- Query 3: Get address
-- Application joins results

Problems

• Multiple round-trips increase latency
• Network overhead multiplied
• No transactional guarantees across queries
• Application complexity for joins
• Inconsistent results if data changes between queries

Solution: Denormalize for Query Patterns

-- Order display query needs: order + user info + shipping address
CREATE TABLE orders_with_details (
    order_id UUID,
    user_id UUID,
    user_name TEXT,           -- Denormalized from users
    user_email TEXT,          -- Denormalized from users
    shipping_street TEXT,     -- Denormalized from addresses
    shipping_city TEXT,
    shipping_country TEXT,
    total DECIMAL,
    created_at TIMESTAMP,
    PRIMARY KEY ((order_id))
);

-- Single query returns everything needed
SELECT * FROM orders_with_details WHERE order_id = ?;

Trade-off Management

Aspect	Trade-off
Storage	More disk space (data duplicated)
Write complexity	Must update multiple tables on change
Consistency	Denormalized data may become stale

Aspect	Benefit
Read performance	Single query, single partition
Latency	Predictable, low latency
Scalability	Horizontal scaling works

---

Anti-Pattern Detection Checklist

Log Analysis

# Tombstone warnings
grep -i "tombstone" /var/log/cassandra/system.log

# Large partition warnings
grep "Compacting large partition" /var/log/cassandra/system.log
grep "Writing large partition" /var/log/cassandra/system.log

# Query warnings
grep "ALLOW FILTERING" /var/log/cassandra/debug.log
grep "Slow query" /var/log/cassandra/debug.log

nodetool Diagnostics

# Partition sizes
nodetool tablehistograms keyspace.table

# Tombstone metrics
nodetool tablestats keyspace.table | grep -i tombstone

# Query latencies
nodetool tablestats keyspace.table | grep -i latency

Query Tracing

TRACING ON;
SELECT * FROM suspicious_table WHERE ...;
-- Analyze trace for:
-- - Tombstones scanned
-- - SSTable count
-- - Node latencies

JMX Metrics

# Key metrics to monitor
ReadLatency, WriteLatency          - Per-table latencies
TombstoneScannedHistogram          - Tombstones per read
EstimatedPartitionSizeHistogram    - Partition size distribution
PendingCompactions                 - Compaction backlog

---

Summary: Design Checklist

Before deploying any Cassandra table:

Check	Requirement
☐ Partition key	Has high cardinality
☐ Partition size	Bounded (< 100 MB)
☐ Growth pattern	No unbounded growth
☐ Queries	Do not require ALLOW FILTERING
☐ Secondary indexes	Not on high-cardinality columns
☐ Collections	Limited to ~100 elements
☐ DELETEs	Minimized (use TTL instead)
☐ Counters	Properly isolated in counter-only tables
☐ Access patterns	Each pattern has a dedicated table
☐ IN clauses	Limited to ~20 values

---

Next Steps

• Data Modeling Concepts - Correct design patterns
• Time Bucketing - Partition management
• Performance Tuning - Optimization
• Troubleshooting - Problem diagnosis