advanced⏱75 minLección 4 de 12

Compute & Processing

Evaluate compute abstractions across platforms. Compare Amazon Redshift's MPP, Snowflake's Virtual Warehouses, and Databricks' Spark clusters with the Photon engine.

Compute & Processing

Bottom Line: Snowflake's Virtual Warehouses win on simplicity and SQL concurrency. Databricks' Spark clusters with Photon win on flexibility and ML workloads. AWS gives you both options — and the complexity of maintaining them.

4.1 Compute Abstraction Levels

Each platform abstracts compute differently. Understanding the abstraction level helps set expectations on what you can — and cannot — control.

💡 Tip: Use Ctrl + Mouse Wheel to zoom, or click the fullscreen icon

4.2 Amazon Redshift

Redshift is AWS's fully managed MPP (Massively Parallel Processing) columnar SQL data warehouse. It is optimised for complex analytical queries on large datasets using SQL.

Redshift Provisioned vs. Serverless

Attribute	Provisioned	Serverless
Billing	Per-hour per node (ra3.xlplus → ra3.16xlarge)	Per RPU-second (min 8 RPUs)
Startup Time	~5 min cluster start	~5 sec (warm)
Concurrency Scaling	Add Concurrency Scaling clusters ($$$)	Automatic
WLM (Workload Management)	Manual queue config	Automatic
AQUA (Accel. QP)	✅ Available on ra3	✅ Available
Best For	Predictable, always-on workloads	Burst/intermittent workloads
Reserved Instance Discount	Up to 75% (1yr/3yr)	❌ No RIs
Max Data (managed storage)	128TB per node (ra3 managed)	Unlimited (Redshift Managed Storage)

Redshift Architecture Internals

                    ┌───────────────────────────────┐
                    │         Leader Node             │
                    │  - SQL parsing & optimisation   │
                    │  - Query coordination           │
                    │  - Result aggregation           │
                    └──────────────┬────────────────┘
                                   │
           ┌───────────────────────┼───────────────────────┐
           │                       │                       │
    ┌──────▼──────┐        ┌───────▼──────┐       ┌───────▼──────┐
    │  Compute    │        │  Compute     │       │  Compute     │
    │  Node 1     │        │  Node 2      │       │  Node N      │
    │  ┌───┐┌───┐ │        │  ┌───┐┌───┐ │       │  ┌───┐┌───┐  │
    │  │S1 ││S2 │ │        │  │S3 ││S4 │ │       │  │SN ││.. │  │
    │  └───┘└───┘ │        │  └───┘└───┘ │       │  └───┘└───┘  │
    │  (Slices)   │        │  (Slices)   │       │  (Slices)    │
    └─────────────┘        └────────────-┘       └──────────────┘
    Each node has local SSD. Data distributed by DISTKEY/DISTSTYLE.

Redshift Distribution Styles — Critical for Performance

sql

-- DISTSTYLE KEY: Distribute rows by a specific column value
-- Use when: Large table joined frequently on a specific column
CREATE TABLE orders (
    order_id    VARCHAR(36),
    customer_id VARCHAR(36),   -- ← Frequently joined with customers table
    revenue     DECIMAL(10,2),
    order_date  DATE
)
DISTKEY(customer_id)    -- Rows with same customer_id go to same node → reduces shuffle
SORTKEY(order_date);    -- Rows stored in order_date order → enables zone map pruning

-- DISTSTYLE ALL: Copy full table to every node
-- Use when: Small dimension tables (< 5M rows) frequently joined to large tables
CREATE TABLE dim_regions (
    region_id   INT,
    region_name VARCHAR(100),
    country     VARCHAR(100)
)
DISTSTYLE ALL;  -- Every node has a full copy — no network shuffle for joins

-- DISTSTYLE EVEN: Round-robin distribution
-- Use when: No clear join key, balanced write throughput is priority
CREATE TABLE staging_events (
    event_id    BIGINT,
    payload     SUPER   -- Semi-structured JSON
)
DISTSTYLE EVEN;

-- ⚠️ GOTCHA: Mismatched DISTKEY between two joined tables = full network shuffle = SLOW

Redshift Spectrum — Querying S3 from Redshift

sql

-- Create an external schema pointing to Glue Catalog
CREATE EXTERNAL SCHEMA spectrum_raw
FROM DATA CATALOG
DATABASE 'raw_ecommerce'
IAM_ROLE 'arn:aws:iam::123456789:role/RedshiftSpectrumRole'
CREATE EXTERNAL DATABASE IF NOT EXISTS;

-- Query S3 data directly (no data loaded into Redshift)
SELECT
    s.order_id,
    s.revenue,
    d.customer_name
FROM spectrum_raw.orders s                  -- Lives in S3
JOIN redshift_local.dim_customers d         -- Lives in Redshift
ON s.customer_id = d.customer_id
WHERE s.order_date >= '2024-01-01';

-- ⚡ Spectrum pushes predicates down to S3 — only matching rows are transferred
-- 💰 Billed at $5 per TB scanned (plus Redshift compute costs)

4.3 Snowflake Virtual Warehouses

Virtual Warehouse Sizing

Snowflake VWs are T-shirt sized. Doubling the size roughly doubles the performance (and cost):

Size	vCores (approx)	Credits/Hour	Use Case
X-Small	1	1	Development, testing
Small	2	2	Light BI, simple queries
Medium	4	4	Standard analytics
Large	8	8	Heavy aggregations
X-Large	16	16	Complex joins, large sorts
2X-Large	32	32	Very large datasets
3X-Large	64	64	Peak load or ML feature engineering
4X-Large	128	128	Snowpark heavy ML workloads
5X-Large	256	256	Extreme workloads
6X-Large	512	512	Largest supported size

💰 Cost Note: Snowflake credits cost approximately $2-4 USD each (varies by edition and cloud region). A 4X-Large running for 1 hour ≈$ 256–512 USD.

Virtual Warehouse Lifecycle Management

sql

-- Create a warehouse with aggressive auto-suspend for cost savings
CREATE WAREHOUSE analytics_wh
    WAREHOUSE_SIZE = 'LARGE'
    AUTO_SUSPEND = 60           -- Suspend after 60 seconds of inactivity
    AUTO_RESUME = TRUE          -- Automatically resume on query
    INITIALLY_SUSPENDED = TRUE; -- Start suspended (save credits)

-- Create a multi-cluster warehouse for high concurrency
CREATE WAREHOUSE reporting_wh
    WAREHOUSE_SIZE = 'MEDIUM'
    MIN_CLUSTER_COUNT = 1
    MAX_CLUSTER_COUNT = 5
    SCALING_POLICY = 'STANDARD' -- 'ECONOMY' waits longer to scale out
    AUTO_SUSPEND = 300;

-- Monitor warehouse credit consumption
SELECT
    warehouse_name,
    start_time,
    end_time,
    credits_used,
    credits_used_cloud_services
FROM SNOWFLAKE.ACCOUNT_USAGE.WAREHOUSE_METERING_HISTORY
WHERE start_time >= DATEADD('day', -7, CURRENT_TIMESTAMP())
ORDER BY credits_used DESC;

Query Acceleration Service (QAS)

Snowflake's Query Acceleration Service offloads portions of eligible queries to serverless compute — particularly useful for large scans that don't benefit from VW size increases:

sql

-- Enable QAS on a warehouse (Enterprise Edition+)
ALTER WAREHOUSE analytics_wh SET
    ENABLE_QUERY_ACCELERATION = TRUE
    QUERY_ACCELERATION_MAX_SCALE_FACTOR = 8;  -- Max 8x the VW's compute

-- QAS automatically identifies and offloads:
-- - Large table scans with selective predicates
-- - Queries with high data spillage
-- - Outlier queries (taking much longer than average)

Snowpark — Beyond SQL

Snowpark extends Snowflake's compute to support Python, Java, and Scala directly in the VW:

python

from snowflake.snowpark import Session
from snowflake.snowpark import functions as F
from snowflake.snowpark.types import IntegerType

# Connect to Snowflake
session = Session.builder.configs({
    "account": "myorg-myaccount",
    "user": "data_engineer",
    "role": "DATA_ENGINEER",
    "warehouse": "ML_WH",
    "database": "ECOMMERCE",
    "schema": "ANALYTICS"
}).create()

# Snowpark DataFrame — executed in Snowflake (not locally)
orders_df = session.table("orders")
customers_df = session.table("dim_customers")

# Complex transformation — all pushed to Snowflake compute
result = (
    orders_df
    .join(customers_df, orders_df["customer_id"] == customers_df["customer_id"])
    .filter(F.col("order_date") >= F.lit("2024-01-01"))
    .with_column("revenue_category",
        F.when(F.col("revenue") > 1000, "high")
         .when(F.col("revenue") > 100, "medium")
         .otherwise("low"))
    .group_by("region", "revenue_category")
    .agg(
        F.sum("revenue").alias("total_revenue"),
        F.count("order_id").alias("order_count")
    )
)

# .show() triggers lazy execution — result runs in Snowflake
result.show()
result.write.mode("overwrite").save_as_table("gold_revenue_by_category")

4.4 Databricks Compute

Cluster Types

┌──────────────────────────────────────────────────────────────┐
│  CLUSTER TYPE          │  USE CASE              │  BILLING   │
├──────────────────────────────────────────────────────────────┤
│  All-Purpose Cluster   │  Interactive notebooks  │  Per-DBU   │
│                        │  Ad-hoc exploration     │  (premium) │
├──────────────────────────────────────────────────────────────┤
│  Job Cluster           │  Automated pipelines    │  Per-DBU   │
│                        │  Scheduled workflows    │  (cheaper) │
├──────────────────────────────────────────────────────────────┤
│  SQL Warehouse         │  SQL analytics / BI     │  Per-DBU   │
│  (Classic / Serverless)│  Dashboards             │  (SQL DBU) │
├──────────────────────────────────────────────────────────────┤
│  DLT Pipeline Cluster  │  Delta Live Tables      │  Per-DBU   │
│                        │  (managed by DLT)       │  (DLT DBU) │
└──────────────────────────────────────────────────────────────┘

Cluster Configuration (PySpark)

python

# Cluster configuration via Databricks REST API or Terraform
cluster_config = {
    "cluster_name": "data-engineering-prod",
    "spark_version": "14.3.x-scala2.12",
    "node_type_id": "r6id.2xlarge",       # Memory-optimised EC2
    "driver_node_type_id": "r6id.4xlarge", # Larger driver for coordination
    "autoscale": {
        "min_workers": 2,
        "max_workers": 20
    },
    "auto_termination_minutes": 30,
    "spark_conf": {
        "spark.sql.adaptive.enabled": "true",       # AQE — critical!
        "spark.sql.adaptive.coalescePartitions.enabled": "true",
        "spark.databricks.delta.optimizeWrite.enabled": "true",
        "spark.databricks.delta.autoCompact.enabled": "true"
    },
    "aws_attributes": {
        "instance_profile_arn": "arn:aws:iam::123456789:instance-profile/databricks-ec2",
        "availability": "SPOT_WITH_FALLBACK",  # 💰 Use spot instances to save ~70%
        "spot_bid_price_percent": 100
    }
}

Photon Engine — Vectorised Query Execution

Photon is Databricks' proprietary C++ query engine that replaces the JVM-based Spark SQL execution for compatible operations:

Without Photon (JVM Spark):
  Query → JVM bytecode → Row-by-row processing → Result
  Speed: Baseline

With Photon (C++ Vectorized):
  Query → LLVM-compiled native code → Column batch processing → Result
  Speed: 2-8x faster for SQL workloads (varies by operation)

Photon-compatible operations: SELECT, GROUP BY, JOIN, SORT, FILTER, window functions, most aggregate functions, Delta Lake reads/writes.

Photon-incompatible (falls back to JVM): Python UDFs, Java UDFs, some complex lambda expressions, certain Spark ML operations.

python

# Check if Photon is being used (from query plan)
df = spark.sql("""
    SELECT customer_id, SUM(revenue) as total
    FROM orders
    WHERE order_date >= '2024-01-01'
    GROUP BY customer_id
""")

# Inspect physical plan — look for PhotonResultStage
df.explain(mode="formatted")
# Output shows: PhotonGroupingAgg, PhotonFilter, PhotonScan

Databricks SQL Warehouses

SQL Warehouses are purpose-built for SQL analytics — separate from Spark clusters, optimised for concurrency:

sql

-- Configure via Databricks UI or API
-- Key settings:
-- Type: Serverless (fastest startup, auto-scale) or Classic (more control)
-- Size: 2X-Small to 4X-Large (similar to Snowflake T-shirt sizing)
-- Auto-stop: 10-480 minutes
-- Scaling: Min 1 cluster, Max N clusters

-- Example: Parameterised SQL query for dashboards
SELECT
    date_trunc('month', order_date) AS month,
    region,
    SUM(revenue)                    AS total_revenue,
    COUNT(DISTINCT customer_id)     AS unique_customers
FROM orders
WHERE
    order_date BETWEEN :start_date AND :end_date  -- Dashboard parameter
    AND region IN (:regions)
GROUP BY 1, 2
ORDER BY 1, 2;

Adaptive Query Execution (AQE) — Databricks' Performance Edge

Spark's AQE re-optimises query plans at runtime based on actual statistics, unlike Snowflake's static planning:

python

# AQE is enabled by default in Databricks Runtime 14+
spark.conf.set("spark.sql.adaptive.enabled", "true")
spark.conf.set("spark.sql.adaptive.coalescePartitions.enabled", "true")
spark.conf.set("spark.sql.adaptive.skewJoin.enabled", "true")

# How AQE helps at runtime:
# 1. Coalesce Partitions: After a shuffle, reduces 10,000 small partitions
#    to 200 optimally-sized partitions automatically
#
# 2. Skew Join: Detects if one partition has 10x more data than others
#    (data skew) and splits it automatically — no manual salting needed
#
# 3. Join Strategy: Switches from SortMergeJoin to BroadcastHashJoin
#    if the runtime size of a table turns out to be small

4.5 Amazon EMR — Managed Spark/Trino

For teams needing full Spark (or Presto/Trino) control on AWS without Databricks:

python

# EMR Serverless — no cluster management, pay per use
import boto3

emr_serverless = boto3.client('emr-serverless', region_name='us-east-1')

# Submit a Spark job
response = emr_serverless.start_job_run(
    applicationId='00f1abc123xyz',
    executionRoleArn='arn:aws:iam::123456789:role/EMRServerlessRole',
    jobDriver={
        'sparkSubmit': {
            'entryPoint': 's3://my-bucket/scripts/process_orders.py',
            'entryPointArguments': ['--date', '2024-03-15'],
            'sparkSubmitParameters': '--conf spark.executor.cores=4 '
                                    '--conf spark.executor.memory=8g '
                                    '--conf spark.driver.cores=2 '
                                    '--conf spark.driver.memory=4g'
        }
    },
    configurationOverrides={
        'monitoringConfiguration': {
            'cloudWatchLoggingConfiguration': {
                'enabled': True,
                'logGroupName': '/emr-serverless/applications'
            }
        }
    }
)

4.6 Compute Comparison Matrix

Dimension	AWS Redshift	AWS EMR	Snowflake VW	Databricks SQL WH	Databricks Clusters
Startup Time	3-5 min (provisioned) / ~5s (serverless)	5-15 min	~2s (warm)	~2s (serverless)	2-10 min
Concurrency	Manual WLM config	Queue-based (YARN)	⚡ Automatic multi-cluster	Auto-scale	Manual or auto-scale
SQL Performance	⚡ Very fast (MPP)	Good (Spark SQL)	⚡ Very fast	⚡ Fast (Photon)	Fast (AQE)
Python/Scala	Limited (UDFs)	⚡ Full Spark	✅ Snowpark	Partial (notebook)	⚡ Full Spark
ML Workloads	❌ Not designed for ML	✅ Good	⚠️ Snowpark ML	⚠️ Limited	⚡ Best-in-class
Auto-Suspend	❌ (Provisioned)	❌ (Long-running)	✅ Configurable	✅ Built-in	✅ Configurable
Spot/Preemptible	❌ Not applicable	✅ (EC2 Spot)	❌ (Managed)	✅ (Spot policies)	✅ (SPOT_WITH_FALLBACK)
Operational Burden	Medium	⚠️ High	⚡ Very Low	Low	Medium

4.7 The 1TB Query Benchmark Perspective

(Illustrative performance benchmarks — actual results vary by query shape, data skew, and configuration)

Workload Type	AWS Redshift (2XL, 2 nodes)	Snowflake (Large VW)	Databricks (Photon, 8-core workers)
Simple point query	<1s	<1s (result cache)	1-3s (cluster startup)
Large aggregation	8-15s	10-20s	12-25s
Multi-table join (5 tables)	15-45s	20-50s	15-40s (AQE)
Full table scan (1TB)	60-120s	45-90s	50-100s
Python ML feature transform	❌ Not applicable	30-90s (Snowpark)	⚡ 10-30s (native Spark)
Streaming aggregation	❌ Not native	❌ Not native	⚡ Native (Structured Streaming)

Progreso33%

Storage & Data Lake

Data Ingestion & Integration

Contenido del Curso

1. Platform Overview & Philosophy 2. Core Architecture Deep Dive 3. Storage & Data Lake 4. Compute & Processing 5. Data Ingestion & Integration 6. Governance, Security & Compliance 7. AI & Machine Learning Capabilities 8. Data Sharing & Collaboration 9. Cost Structure & Optimization 10. Use Case Decision Framework 11. Migration Strategies 12. Reference Architecture Patterns