Amazon Redshift Guide

Redshift is NOT PostgreSQL (read first)

Redshift speaks PostgreSQL's wire protocol and shares much of its surface syntax, so LLMs assume PostgreSQL behavior carries over — it frequently does not. Divergences span system tables (pg_catalog is incomplete), DDL (no indexes, no sequences), functions (string_agg, SUBSTR on tables, leader-node-only functions), types (a text column becomes VARCHAR(256)), and comparison semantics (trailing blanks, unenforced constraints). Assume divergence and verify against the reference below — do not answer from PostgreSQL habit. Common PostgreSQL→Redshift divergences are in references/redshift-sql-syntax.md.

Works best with the AWS MCP server — it runs the AWS CLI and Redshift Data API calls below in a sandboxed, audit-logged environment. All guidance here is plain AWS CLI and SQL and works without it.

STEP 0: Serverless or Provisioned?

Establish this before answering — APIs, system tables, and capabilities differ. Take it from the question when it says which one; ask when it does not. SELECT version() does not identify it.

• Serverless — identified by a workgroup (and namespace). Data API calls take --workgroup-name; the user says "workgroup"/"Serverless".
• Provisioned — identified by a cluster. Data API calls take --cluster-identifier; the user says "cluster".

Critical Facts

• SHOW commands are the primary metadata interface — SHOW DATABASES, SHOW SCHEMAS, SHOW TABLES, SHOW COLUMNS, SHOW TABLE, SHOW VIEW. Do NOT default to pg_catalog or information_schema. → Load references/redshift-sql-metadata.md for metadata/discovery questions and any "relation does not exist" report — it has the diagnostic flow.
• SYS_ views are the preferred system views — they work everywhere. STL_, STV_, SVL_, and SVCS_ are provisioned single-AZ only, and some SVV_ views are unsupported on Serverless. → Load references/redshift-sql-metadata.md for any system-view or monitoring question.
• sys_load_error_detail for COPY debugging (not stl_load_errors, which is provisioned single-AZ only).
• DATEADD/DATEDIFF — unit-first argument order: DATEADD(day, -30, GETDATE()), DATEDIFF(day, start, end).
• APPROXIMATE COUNT(DISTINCT col) — Redshift-specific, ~2% error, much faster than exact COUNT(DISTINCT) on large datasets.
• MERGE ... REMOVE DUPLICATES — simplified dedup when source and target have identical schemas.
• COPY should use IAM_ROLE (the namespace role, not the caller role) + supports MANIFEST for explicit file lists + MAXERROR for error tolerance.
• SUBSTR() is leader-node-only — works on literals but errors on table columns (SUBSTR() function is not supported (Hint: use SUBSTRING instead)). Use SUBSTRING() on columns.
• UNIQUE / PRIMARY KEY / FOREIGN KEY are informational only — NOT enforced (duplicate rows are accepted with no error). Optimizer hints; enforce integrity in the application or via MERGE. NOT NULL IS enforced.
• SHOW VIEW <schema.name> returns the definition of a regular view, materialized view, or late-binding view. MV freshness: SVV_MV_INFO (is_stale).
• No FILTER (WHERE ...) on aggregates — PostgreSQL-only, Redshift rejects it. Write COUNT(CASE WHEN cond THEN 1 END) / SUM(CASE WHEN cond THEN x END) instead. This applies to every query you emit, including ones you derive from a reference example.
• TOP N and LIMIT N both work (TOP N PERCENT does not). A text column becomes VARCHAR(256) — use VARCHAR(max) or explicit length.
• Iceberg tables use CREATE TABLE ... USING ICEBERG (not STORED AS ICEBERG, not TABLE_FORMAT=ICEBERG).
• Datashares support read and write operations — consumers can write once the producer grants write privileges. Treat "permission denied" on a datashare write as a missing grant, not an unsupported operation. → Load references/redshift-sql-metadata.md for requirements and limits.
• Storage: attribute before you remediate, and never reflex-VACUUM. Disk-full is usually NOT data growth — rule out transient query spill (temp blocks) and per-node/slice skew first, because the cluster average hides a single full node. On RA3/Serverless a high local reading is cache, not out-of-space, and free managed storage does not prevent a node-local Disk Full. SVV_TABLE_INFO.unsorted alone is misleading — gate sorting on vacuum_sort_benefit. "Storage is growing / what is driving it" is a different question from "storage is high": it needs a trend over months (CloudWatch, not a point-in-time view), per-table attribution is always partial, and the per-table sum is never the cluster total. → Load references/redshift-storage-utilization.md to attribute space, then references/redshift-storage-remediation.md to act.

Safety Guardrails

BLOCK: DROP DATABASE, DELETE without WHERE, publicly-accessible=true, GRANT ALL ON ALL WARN then confirm: RESIZE, RESTORE, VACUUM on large tables, ALTER PASSWORD, WLM config change, TRUNCATE / DROP TABLE — TRUNCATE forces an implicit COMMIT and cannot be rolled back (it also commits other pending work in the same transaction), so name the target table and say the rows are unrecoverable before proceeding; treat it as routine only on temp/staging tables the user has already called disposable Confirm: CREATE, GRANT specific, COPY, UNLOAD

Security Considerations

Apply these defaults when generating anything that connects, loads, or exports. Details are in the reference files noted.

• In transit: the Data API is HTTPS-only. For JDBC/ODBC set the require_ssl parameter and connect with sslmode=verify-full so the server certificate is checked.
• At rest: keep cluster/namespace encryption enabled, and add ENCRYPTED KMS_KEY_ID '<arn>' to UNLOAD — it writes query results to S3, outside Redshift's own encryption. → references/redshift-sql-ddl-copy.md
• Credentials: prefer SecretArn (Secrets Manager) or IAM Identity Center; DbUser is acceptable because it issues temporary credentials. Never place database passwords in code, environment variables, or SQL text. → references/redshift-sql-recipes-load-api.md
• Least privilege: scope the namespace IAM_ROLE to the specific bucket and prefix (s3:GetObject on arn:aws:s3:::<bucket>/<prefix>/*), not s3:* or a managed full-access policy, and condition its trust policy on both aws:SourceArn (the cluster/namespace ARN) and aws:SourceAccount — SourceArn alone still allows another resource in the account to assume it. Grant per-object privileges rather than GRANT ALL ON ALL.
• Audit: CloudTrail records redshift-data:* API calls but not the SQL executed; enable Redshift audit logging (useractivitylog, connectionlog, userlog) for that. Both capture query text and user activity, so encrypt every destination in use: the CloudWatch Logs group (aws logs associate-kms-key), the CloudTrail trail (SSE-KMS), and the audit-log S3 bucket (SSE-S3 — audit logging to S3 supports only S3-managed keys, not KMS). Serverless only supports sending audit logs to CloudWatch.
• Network: keep PubliclyAccessible=false and connect over a VPC endpoint. Do not open port 5439 to 0.0.0.0/0 or ::/0 — scope inbound rules to specific CIDRs or to a referencing security group.
• Sensitive data: Data API results persist for 24h and sys_load_error_detail can echo fragments of rejected rows, so treat statement IDs and load-error output as sensitive.
• Further reading: Security in Amazon Redshift for the full guidance behind these defaults.

Routing Table

MANDATORY: When a question matches a row below, you MUST load and read the referenced file BEFORE answering.

Ask whether the target is provisioned or Serverless before giving troubleshooting steps — unless the question already says which one, in which case use that and do not re-confirm.

Data API Quick Reference

→ Load references/redshift-sql-recipes-load-api.md before answering ANY Data API, COPY-error, or async-query question. It carries the bounded poll loop, the HasResultSet and ResourceNotFoundException handling, the per-target parameters, and the auth options.

Data API calls are async by default — use long polling (--wait-time-seconds, 1–30) rather than blind sleeps, and keep a bounded loop for work that can exceed 30s. Serverless takes --workgroup-name, provisioned takes --cluster-identifier.