advanced⏱60 minLección 3 de 12

Storage & Data Lake

Learn how each platform manages cloud object storage. Compare AWS S3 with Lake Formation, Snowflake's proprietary micro-partitions, and Databricks' Delta Lake transaction log.

Storage & Data Lake

Bottom Line: S3 is the universal substrate that all three platforms can use, but each platform adds a different reliability and management layer on top. Snowflake abstracts it away entirely; Databricks makes it first-class with Delta Lake; AWS gives you the raw controls and the governance tools to manage it yourself.

3.1 The Object Storage Foundation

All three platforms ultimately store data in cloud object storage (S3, ADLS Gen2, or GCS). The differentiation is the management and reliability layer built on top.

                    What each platform does with Object Storage:

  AWS              ████████████████████████████████████████████████
  (You manage)     S3  →  [YOU define: folders, formats, partitions, compaction]
                           Governance via: Lake Formation + Glue Catalog

  Snowflake        ████████████████████████████████████████████████
  (They manage)    S3  →  [Snowflake manages: FDN format, micro-partitions, clustering]
                           Your data is ON S3, but you can't directly read it (proprietary format)

  Databricks       ████████████████████████████████████████████████
  (Shared)         S3  →  Delta Log (JSON) + Parquet files
                           YOUR S3, YOUR format, YOU can read it with any Spark engine

3.2 Amazon S3 as a Data Lake (AWS)

Core S3 Concepts for Data Platforms

S3 is not "just storage" — it has evolved into a sophisticated foundation for data lakes:

s3://my-data-lake/
├── raw/                          ← Bronze zone (untouched source data)
│   ├── crm/salesforce/year=2024/month=01/day=01/
│   │   └── contacts_20240101.json.gz
│   └── ecommerce/orders/year=2024/month=01/
│       └── orders_part_00001.parquet
├── curated/                      ← Silver zone (cleaned, typed, joined)
│   └── customer_360/
│       └── year=2024/month=01/
│           └── *.snappy.parquet  ← Hive-style partitioning
└── aggregated/                   ← Gold zone (business-level aggregates)
    └── regional_revenue/
        └── *.parquet

S3 Storage Classes and Cost Optimization

Storage Class	Use Case	Retrieval Time	Relative Cost
S3 Standard	Hot data, active queries	Immediate	$$$
S3 Intelligent-Tiering	Unknown access patterns	Immediate (Standard tier)	$$ + monitoring fee
S3 Standard-IA	Infrequent access (>1 month)	Immediate	$$
S3 Glacier Instant Retrieval	Archive, quarterly access	Immediate	$
S3 Glacier Deep Archive	Compliance cold storage	12 hrs	¢

💰 Cost Gotcha: Athena and Glue jobs are billed on data scanned, not stored. A poorly partitioned table in S3 Standard can cost 10x more to query than a well-partitioned table in S3 IA, even if the storage cost is similar.

AWS Glue Data Catalog

The Glue Catalog is the central metadata registry for AWS data lake assets. It is Hive Metastore-compatible and used by Athena, EMR, Redshift Spectrum, and Glue ETL jobs.

python

# Create a table in Glue Catalog using Boto3
import boto3

glue = boto3.client('glue')

glue.create_table(
    DatabaseName='ecommerce',
    TableInput={
        'Name': 'orders',
        'StorageDescriptor': {
            'Columns': [
                {'Name': 'order_id', 'Type': 'string'},
                {'Name': 'customer_id', 'Type': 'string'},
                {'Name': 'revenue', 'Type': 'double'},
                {'Name': 'order_date', 'Type': 'date'}
            ],
            'Location': 's3://my-data-lake/curated/orders/',
            'InputFormat': 'org.apache.hadoop.mapred.TextInputFormat',
            'OutputFormat': 'org.apache.hadoop.hive.ql.io.HiveIgnoreKeyTextOutputFormat',
            'SerdeInfo': {
                'SerializationLibrary': 'org.apache.hadoop.hive.ql.io.parquet.serde.ParquetHiveSerDe'
            }
        },
        'PartitionKeys': [
            {'Name': 'year', 'Type': 'string'},
            {'Name': 'month', 'Type': 'string'}
        ],
        'TableType': 'EXTERNAL_TABLE'
    }
)

AWS Lake Formation: Adding Governance to S3

Lake Formation sits above S3 and Glue Catalog to provide fine-grained access control. Without Lake Formation, S3 access is all-or-nothing via IAM bucket policies.

WITHOUT Lake Formation:
  IAM Policy: s3:GetObject on s3://data-lake/*
  → User can read ALL files in the bucket (no column/row-level control)

WITH Lake Formation:
  LF Permission: SELECT on database=ecommerce, table=orders, columns=[order_id, revenue]
  → User can only SELECT two specific columns, even if they have S3 access
  
  LF Row Filter: revenue > 0 AND region = 'APAC'
  → User only sees rows matching the filter

Lake Formation Table Formats:

Lake Formation supports governed tables with ACID properties:

sql

-- Create a Lake Formation Governed Table (ACID on S3)
CREATE TABLE ecommerce.orders
  LOCATION 's3://my-data-lake/governed/orders/'
  TBLPROPERTIES (
    'table_type' = 'GOVERNED'
  );

-- ACID inserts via Athena (Lake Formation handles the transaction log)
INSERT INTO ecommerce.orders VALUES
  ('ord-001', 'cust-001', 299.99, DATE '2024-01-15');

⚠️ Limitation: Lake Formation Governed Tables use AWS's own transaction log format, which is NOT compatible with Delta Lake or Apache Iceberg. Prefer Apache Iceberg on S3 if you need open-format ACID tables on AWS.

Apache Iceberg on AWS (The Open Format Recommendation)

For organisations wanting open-format ACID transactions on AWS (comparable to Delta Lake), Apache Iceberg via Athena or EMR is the recommended path:

sql

-- Create an Iceberg table in Athena
CREATE TABLE ecommerce.orders (
    order_id      VARCHAR,
    customer_id   VARCHAR,
    revenue       DOUBLE,
    order_date    DATE
)
PARTITIONED BY (year(order_date))
LOCATION 's3://my-data-lake/iceberg/orders/'
TBLPROPERTIES (
    'table_type' = 'ICEBERG',
    'write_compression' = 'snappy',
    'optimize_rewrite_delete_file_threshold' = '10'
);

-- MERGE (upsert) — fully ACID
MERGE INTO ecommerce.orders t
USING (SELECT * FROM staging.orders_updates) s
ON t.order_id = s.order_id
WHEN MATCHED THEN UPDATE SET revenue = s.revenue
WHEN NOT MATCHED THEN INSERT *;

3.3 Snowflake Storage

How Snowflake Manages Storage

Snowflake abstracts storage completely. The "database" you interact with via SQL is backed by Snowflake's internal File/Data Node (FDN) format, stored in Snowflake's own S3/GCS/Azure Blob tenant. Key properties:

Immutable micro-partitions: 50–500MB compressed blocks, automatically created on INSERT.
Automatic clustering: Snowflake tracks the range of values in each column per micro-partition, enabling automatic partition pruning without explicit indexes.
Columnar compression: Each column within a micro-partition is compressed independently using the optimal algorithm (LZO, Zstd, etc.).

sql

-- You interact with Snowflake storage purely via SQL
-- All physical layout decisions are automatic

CREATE TABLE orders (
    order_id      VARCHAR,
    customer_id   VARCHAR,
    revenue       NUMBER(10,2),
    order_date    DATE,
    region        VARCHAR
)
-- No SORTKEY, DISTKEY, PARTITION BY required (unlike Redshift)
-- Snowflake handles physical layout automatically
;

-- Optional: Add a clustering key for very large tables (>1TB)
-- where access patterns are highly selective on a specific column
ALTER TABLE orders CLUSTER BY (order_date);
-- Snowflake will automatically re-cluster in the background (costs credits 💰)

Snowflake External Tables and Stages

Snowflake can also query data directly from S3/GCS/ADLS without loading it — useful for landing zones:

sql

-- Step 1: Create a Storage Integration (one-time, admin task)
CREATE STORAGE INTEGRATION s3_data_lake_int
    TYPE = EXTERNAL_STAGE
    STORAGE_PROVIDER = 'S3'
    ENABLED = TRUE
    STORAGE_AWS_ROLE_ARN = 'arn:aws:iam::123456789:role/snowflake-s3-role'
    STORAGE_ALLOWED_LOCATIONS = ('s3://my-data-lake/');

-- Step 2: Create a Named Stage
CREATE STAGE raw_orders_stage
    STORAGE_INTEGRATION = s3_data_lake_int
    URL = 's3://my-data-lake/raw/orders/'
    FILE_FORMAT = (TYPE = 'PARQUET');

-- Step 3: Create an External Table (queries S3 directly, no data copy)
CREATE EXTERNAL TABLE ext_orders (
    order_id      VARCHAR AS (VALUE:order_id::VARCHAR),
    customer_id   VARCHAR AS (VALUE:customer_id::VARCHAR),
    revenue       FLOAT   AS (VALUE:revenue::FLOAT),
    order_date    DATE    AS (VALUE:order_date::DATE)
)
PARTITION BY (order_date)
LOCATION = @raw_orders_stage
AUTO_REFRESH = TRUE  -- Snowflake monitors S3 event notifications
FILE_FORMAT = (TYPE = 'PARQUET');

Snowflake Time Travel and Fail-Safe

Feature	Time Travel	Fail-Safe
Purpose	User-accessible history (queries, clone, undrop)	Disaster recovery (Snowflake-managed)
Duration	0–90 days (Standard=1, Enterprise=90)	Fixed 7 days (after Time Travel period)
Accessible by	You (via AT/BEFORE clause)	Snowflake Support only
Cost	✅ Included in storage cost	✅ Included in storage cost

sql

-- Query data as it was 2 days ago
SELECT * FROM orders AT (OFFSET => -172800);  -- seconds

-- Query at a specific timestamp
SELECT * FROM orders AT (TIMESTAMP => '2024-03-15 09:00:00'::TIMESTAMP_LTZ);

-- Restore a dropped table
UNDROP TABLE orders;

-- Zero-Copy Clone (reference original micro-partitions — no data copy)
CREATE TABLE orders_dev CLONE orders;  -- Instant, storage cost = 0 until writes diverge

3.4 Delta Lake on Databricks

The Delta Lake Transaction Log in Detail

Delta Lake's durability and reliability come from the _delta_log/ directory. Understanding this is essential for operational work:

s3://my-bucket/delta-table/
├── _delta_log/
│   ├── 00000000000000000000.json   {"commitInfo":..., "protocol":..., "metaData":...}
│   ├── 00000000000000000001.json   {"add": {"path": "part-001.parquet", ...}}
│   ├── 00000000000000000002.json   {"remove": {"path": "part-001.parquet",...},
│   │                                "add": {"path": "part-003.parquet",...}}
│   └── 00000000000000000010.checkpoint.parquet  ← Snapshot at commit 10
├── part-00001-uuid.snappy.parquet  ← Still physically exists (soft-deleted)
├── part-00002-uuid.snappy.parquet
└── part-00003-uuid.snappy.parquet  ← Newest data file

This log enables:

Snapshot isolation: Readers see a consistent version; writers don't block readers.
Time Travel: Read old snapshots by replaying the log to a target version.
Audit History: Every change is recorded with timestamp, user, and operation.

python

# Read a specific Delta table version (Time Travel)
df = spark.read.format("delta") \
    .option("versionAsOf", 5) \
    .load("s3://my-bucket/delta-table/")

# Or by timestamp
df = spark.read.format("delta") \
    .option("timestampAsOf", "2024-03-15") \
    .load("s3://my-bucket/delta-table/")

# SQL equivalents
spark.sql("SELECT * FROM orders VERSION AS OF 5")
spark.sql("SELECT * FROM orders TIMESTAMP AS OF '2024-03-15'")

Delta Lake: Key Operations

python

from delta.tables import DeltaTable
from pyspark.sql import functions as F

# --- MERGE (Upsert) ---
deltaTable = DeltaTable.forPath(spark, "s3://my-bucket/delta/orders/")
updates_df = spark.read.parquet("s3://my-bucket/raw/order_updates/")

deltaTable.alias("t").merge(
    updates_df.alias("s"),
    "t.order_id = s.order_id"
).whenMatchedUpdate(set={
    "revenue": "s.revenue",
    "status": "s.status",
    "updated_at": F.current_timestamp()
}).whenNotMatchedInsert(values={
    "order_id": "s.order_id",
    "revenue": "s.revenue",
    "status": "s.status",
    "created_at": F.current_timestamp(),
    "updated_at": F.current_timestamp()
}).execute()

# --- OPTIMIZE (Compaction + Z-Ordering) ---
# Compact small files AND co-locate related data for faster filtering
spark.sql("""
    OPTIMIZE delta.`s3://my-bucket/delta/orders/`
    ZORDER BY (customer_id, order_date)
""")
# ZORDER co-locates customer_id + order_date values in the same files
# Dramatically improves queries that filter on both columns

# --- VACUUM (Remove old files no longer referenced) ---
spark.sql("""
    VACUUM delta.`s3://my-bucket/delta/orders/`
    RETAIN 168 HOURS  -- 7 days minimum recommended
""")
# WARNING: After VACUUM, Time Travel to pre-VACUUM versions is no longer possible

Databricks Auto Optimize Features

Databricks adds managed optimisation on top of open-source Delta Lake:

sql

-- Enable Auto Optimize at table creation
CREATE TABLE orders (
    order_id    STRING,
    customer_id STRING,
    revenue     DOUBLE,
    order_date  DATE
)
USING DELTA
TBLPROPERTIES (
    'delta.autoOptimize.optimizeWrite' = 'true',  -- Compacts during writes
    'delta.autoOptimize.autoCompact'   = 'true',  -- Async compaction post-write
    'delta.dataSkippingNumIndexedCols' = '4'       -- Stats collected on first 4 cols
);

3.5 Medallion Architecture: Implementation per Platform

The Medallion (Bronze/Silver/Gold) pattern is a universal data lake design pattern. Here's how it maps to each platform:

Layer	Purpose	AWS Implementation	Snowflake Implementation	Databricks Implementation
Bronze	Raw, immutable ingestion	S3 prefix `/raw/`, Glue Catalog, no transformations	Raw staging tables or External Tables on S3	Delta tables with `PERMISSIVE` mode, schema-on-read
Silver	Cleaned, conformed, joined	S3 `/curated/`, Glue ETL or EMR Spark jobs, Parquet	Snowflake tables after Snowpipe/COPY INTO	Delta tables after DLT (Delta Live Tables) validation
Gold	Business-level aggregates	S3 `/aggregated/`, served by Athena or Redshift Spectrum	Snowflake views + aggregation tables	Delta tables, served by SQL Warehouses or Databricks dashboards

Databricks DLT (Delta Live Tables) — Declarative Medallion:

import dlt
from pyspark.sql import functions as F

# Bronze — raw ingestion
@dlt.table(comment="Raw orders from source systems")
def bronze_orders():
    return (
        spark.readStream.format("cloudFiles")
            .option("cloudFiles.format", "json")
            .load("s3://my-bucket/landing/orders/")
    )

# Silver — validated and cleaned
@dlt.table(comment="Validated orders with quality checks")
@dlt.expect_or_drop("valid_revenue", "revenue > 0")
@dlt.expect_or_drop("valid_order_id", "order_id IS NOT NULL")
def silver_orders():
    return (
        dlt.read_stream("bronze_orders")
            .withColumn("order_date", F.to_date("order_date_str", "yyyy-MM-dd"))
            .withColumn("revenue", F.col("revenue").cast("double"))
            .drop("order_date_str")
    )

# Gold — business aggregate
@dlt.table(comment="Daily revenue by region")
def gold_daily_revenue():
    return (
        dlt.read("silver_orders")
            .groupBy("region", "order_date")
            .agg(F.sum("revenue").alias("total_revenue"),
                 F.count("order_id").alias("order_count"))
    )

3.6 Storage Feature Comparison Matrix

Feature	AWS S3 + Lake Formation	Snowflake	Databricks Delta Lake
ACID Transactions	⚠️ Via Iceberg or Governed Tables	✅ Native	✅ Native
Time Travel	⚠️ Via S3 Versioning (costly) or Iceberg	✅ 0–90 days	✅ 30 days (default)
Zero-Copy Clone	❌ Not native	⚡ Instant clone	⚠️ Shallow clone (metadata only)
Auto Compaction	❌ Manual Glue/EMR job	✅ Automatic	✅ Auto Optimize (Databricks)
Schema Evolution	⚠️ Manual Glue update	✅ Automatic with policy	✅ `mergeSchema` option
Row-Level Security	✅ LF Row Filters	✅ Row Access Policies	✅ Unity Catalog row filters
Column Masking	✅ LF Column Masking	✅ Dynamic Data Masking	✅ Unity Catalog column masks
Open Format	✅ Parquet/ORC/Iceberg	❌ Proprietary (internal)	✅ Parquet + Delta spec
Storage Cost Model	Pay per GB + requests	Pay per TB compressed	Pay per GB (your own S3 account)
Multi-Cloud Storage	❌ AWS only	✅ S3/GCS/ADLS	✅ S3/GCS/ADLS

Progreso25%

Core Architecture Deep Dive

Compute & Processing

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