Snowpipe Streaming — Python SDK

This skill is self-contained. All source code, configs, and dependencies are bundled. The agent's job is to copy, configure, and run — not to generate code from scratch.

When to Use

• User wants to stream data into Snowflake in real-time using Python
• User mentions Snowpipe Streaming, the Python SDK, or real-time ingestion
• User needs to set up RSA key-pair auth for streaming
• User wants to run single or parallel streaming instances
• User needs to troubleshoot JWT, key, or permission errors with streaming

Tools Used

• bash — Copy files, run setup.sh, execute streaming scripts
• snowflake_sql_execute — Create databases, schemas, tables (when CoCo connection matches target)
• ask_user_question — Confirm target objects, get approval at checkpoints
• read / write / edit — Configure profile.json and config.properties

Bundled Files

snowpipe-streaming-python/
├── SKILL.md                        # This file (agent instructions)
├── LEARNINGS.md                    # Append-only log of run deviations and fixes
├── FEEDBACK_TEMPLATE.md            # Template for new learning entries
├── requirements.txt                # Python dependencies
├── .gitignore                      # Protects keys and credentials
├── setup.sh                        # Bootstrap: venv + deps + keys + config templates
├── profile.template.json           # Snowflake connection template
├── config.template.properties      # Streaming config template
└── src/
    ├── models.py                   # Customer, Order, OrderItem dataclasses
    ├── data_generator.py           # Synthetic e-commerce data (customer IDs 1-10000)
    ├── config_manager.py           # Loads .properties + profile JSON, resolves key files
    ├── snowpipe_streaming_manager.py  # SDK client/channel lifecycle + retry logic
    ├── reconciliation_manager.py   # Post-ingestion row count + orphan check
    ├── streaming_app.py            # Single-instance entry point
    └── parallel_streaming_orchestrator.py  # Multi-instance entry point

The skill base directory is available via the skill loader. Reference it as SKILL_DIR in instructions below.

Stopping Points

• ✋ Phase 0: User approves the workflow before any action
• ✋ Step 1: User confirms Snowflake target objects (database, schema, tables, role)
• ✋ Step 3: RSA public key registered with Snowflake user (if new key generated)
• ✋ Step 4: profile.json correctly configured before proceeding
• ✋ Step 7: Confirm data landed in Snowflake

Phase 0: Briefing & Consent

Goal: Explain what this skill does and get explicit user approval before executing anything.

⚠️ STOP: This phase MUST be completed before ANY other action. Do not run any commands, read any files, or execute any tools until the user approves.

Present the following briefing to the user:

Snowpipe Streaming (Python) — What This Skill Does

Snowpipe Streaming is Snowflake's real-time data ingestion API. Instead of staging files and loading them in batches, it streams rows directly into tables with sub-second latency. This skill uses the Python SDK (snowflake.ingest.streaming, v1.2.0+) backed by a Rust engine.

The skill is self-contained — all source code, configs, and dependencies are bundled. The agent's job is to copy, configure, and run — not generate code from scratch.

What will happen (7 steps):

1. Gather target details — confirm which Snowflake account, database, schema, tables, and role to use
2. Copy project files — copy bundled source code, configs, and requirements to a project directory
3. Bootstrap environment — run setup.sh to create venv, install deps, generate RSA keys
4. Configure — fill in profile.json and config.properties with your Snowflake details
5. Create Snowflake objects — create the target database, schema, and tables (owned by your role)
6. Run streaming — stream synthetic e-commerce orders + order items into Snowflake
7. Verify data — query row counts after ingestion completes

What this skill will NOT do:

• Drop or alter any existing objects not created by this workflow
• Modify files outside the project directory
• Run anything without showing you the exact command first
• Continue past any checkpoint without your explicit approval

Authentication: Uses RSA key-pair auth (not password). A new key pair will be generated and the public key registered with your Snowflake user via ALTER USER.

Ask the user: "Shall I proceed with Step 1 (Gather Target Details)?"

⚠️ STOP: Do NOT proceed until the user confirms.

Workflow

Pre-flight: Before starting, read LEARNINGS.md in the skill directory. If there are PENDING entries, apply fixes to bundled files before copying them to the user's project. Update applied entries from PENDING to APPLIED.

Step 1: Propose Target Objects

Confirm the Snowflake objects with the user. Propose defaults and let them override:

⚠️ STOP: Ask the user to confirm or change these before proceeding.

Table ownership: The streaming SDK's role must own the target database, schema, and tables. GRANT USAGE or GRANT ALL is not sufficient — the SDK's internal pipe-info API returns ERR_TABLE_DOES_NOT_EXIST_NOT_AUTHORIZED if the role doesn't own the objects.

Step 2: Copy Project Files

Copy the entire skill directory to the user's desired project location:

SKILL_DIR="<path to snowpipe-streaming-python skill>"
PROJECT_DIR="<user's chosen project directory>"

mkdir -p "${PROJECT_DIR}"
cp -r "${SKILL_DIR}/src" "${PROJECT_DIR}/src"
cp "${SKILL_DIR}/requirements.txt" "${PROJECT_DIR}/"
cp "${SKILL_DIR}/setup.sh" "${PROJECT_DIR}/"
cp "${SKILL_DIR}/profile.template.json" "${PROJECT_DIR}/"
cp "${SKILL_DIR}/config.template.properties" "${PROJECT_DIR}/"
cp "${SKILL_DIR}/.gitignore" "${PROJECT_DIR}/"

Do not copy SKILL.md — it's agent instructions, not user-facing.

Step 3: Bootstrap Environment

Run the setup script:

cd "${PROJECT_DIR}"
bash setup.sh

This will:

1. Create .venv/ with a Python virtual environment
2. Install all packages from requirements.txt
3. Generate streaming_key.p8 and streaming_key.pub (if not present)
4. Copy template configs to profile.json and config.properties (if not present)

If the user already has an RSA key (e.g., rsa_key.p8), copy or symlink it as streaming_key.p8 before running setup, or update profile.json to point to the existing key path.

If error occurs:

• pip install failure → check Python version (3.8+ required) and network connectivity
• Key generation failure → ensure openssl is available on PATH

⚠️ STOP: If setup.sh generated a new key pair, the public key must be registered with the Snowflake user before continuing:

ALTER USER <username> SET RSA_PUBLIC_KEY='<contents of streaming_key.pub without header/footer>';

Cross-account key registration: If the target is a different account from the CoCo connection, either:

1. Add a key-pair auth connection to the target account and use it
2. Have the user run ALTER USER directly in Snowsight for the target account

Step 4: Configure Profile and Properties

Edit profile.json with the user's confirmed values from Step 1:

{
  "user": "<USER>",
  "account": "<ORG>-<ACCOUNT>",
  "url": "https://<ORG>-<ACCOUNT>.snowflakecomputing.com:443",
  "role": "<ROLE>",
  "private_key_file": "streaming_key.p8",
  "database": "<DATABASE>",
  "schema": "<SCHEMA>",
  "warehouse": "<WAREHOUSE>"
}

Edit config.properties if the user changed table names from defaults — update the pipe names:

pipe.orders.name=<ORDERS_TABLE>-STREAMING
pipe.order_items.name=<ORDER_ITEMS_TABLE>-STREAMING

Use private_key_file (path to .p8 file) instead of inline private_key. The bundled config_manager.py reads the file at runtime and passes the full PEM to the SDK.

⚠️ STOP: Confirm profile.json is correctly filled in before continuing.

Step 5: Create Snowflake Objects

CoCo connection vs streaming profile: The Cortex Code SQL connection may route to a different account or user than the streaming target. If so, you cannot create tables via the CoCo SQL tool. Instead, use a Python script that connects via the streaming RSA key.

If CoCo connection matches the streaming target account + role, use SQL directly:

CREATE DATABASE IF NOT EXISTS <DATABASE>;
CREATE SCHEMA IF NOT EXISTS <DATABASE>.<SCHEMA>;
CREATE TABLE IF NOT EXISTS <DATABASE>.<SCHEMA>.ORDERS (ORDER_ID VARCHAR, CUSTOMER_ID VARCHAR, ORDER_DATE TIMESTAMP_NTZ, STATUS VARCHAR, TOTAL_AMOUNT NUMBER(12,2));
CREATE TABLE IF NOT EXISTS <DATABASE>.<SCHEMA>.ORDER_ITEMS (ORDER_ITEM_ID VARCHAR, ORDER_ID VARCHAR, PRODUCT_NAME VARCHAR, CATEGORY VARCHAR, QUANTITY NUMBER, UNIT_PRICE NUMBER(12,2), TOTAL_PRICE NUMBER(12,2));

If CoCo connection differs, write and run a temporary Python setup script using the project's venv:

import snowflake.connector
from cryptography.hazmat.primitives import serialization
from cryptography.hazmat.backends import default_backend

with open("streaming_key.p8", "rb") as f:
    p_key = serialization.load_pem_private_key(f.read(), password=None, backend=default_backend())
pkb = p_key.private_bytes(encoding=serialization.Encoding.DER, format=serialization.PrivateFormat.PKCS8, encryption_algorithm=serialization.NoEncryption())

conn = snowflake.connector.connect(user="<USER>", account="<ACCOUNT>", private_key=pkb, role="<ROLE>", warehouse="<WAREHOUSE>")
cursor = conn.cursor()
for stmt in [
    "CREATE DATABASE IF NOT EXISTS <DATABASE>",
    "CREATE SCHEMA IF NOT EXISTS <DATABASE>.<SCHEMA>",
    "CREATE TABLE IF NOT EXISTS <DATABASE>.<SCHEMA>.ORDERS (ORDER_ID VARCHAR, CUSTOMER_ID VARCHAR, ORDER_DATE TIMESTAMP_NTZ, STATUS VARCHAR, TOTAL_AMOUNT NUMBER(12,2))",
    "CREATE TABLE IF NOT EXISTS <DATABASE>.<SCHEMA>.ORDER_ITEMS (ORDER_ITEM_ID VARCHAR, ORDER_ID VARCHAR, PRODUCT_NAME VARCHAR, CATEGORY VARCHAR, QUANTITY NUMBER, UNIT_PRICE NUMBER(12,2), TOTAL_PRICE NUMBER(12,2))",
]:
    cursor.execute(stmt)
    print(cursor.fetchone()[0])
cursor.close(); conn.close()

Streaming pipes are auto-created by Snowflake with the naming convention <TABLE_NAME>-STREAMING.

Step 6: Run Streaming

Customer IDs are generated synthetically (1-10000). No database lookup or CUSTOMERS table needed.

All commands use the project venv Python: .venv/bin/python

Single instance (1K-50K orders):

.venv/bin/python src/streaming_app.py <num_orders> [config_file] [profile_file]

Parallel (10K+ orders, recommended):

.venv/bin/python src/parallel_streaming_orchestrator.py <total_orders> <num_instances> [config_file] [profile_file]

Examples:

# 1K orders, single instance
.venv/bin/python src/streaming_app.py 1000

# 10K orders, 5 parallel instances
.venv/bin/python src/parallel_streaming_orchestrator.py 10000 5

# 100K orders, 10 parallel instances, staging config
.venv/bin/python src/parallel_streaming_orchestrator.py 100000 10 config_staging.properties profile_staging.json

Use run_in_background=true in the Bash tool for long runs.

If error occurs:

• ModuleNotFoundError → use .venv/bin/python, not system Python
• ERR_TABLE_DOES_NOT_EXIST_NOT_AUTHORIZED → role doesn't own tables (see Troubleshooting)
• JWT / auth errors → check profile.json account format and key registration
• ReceiverSaturated / HTTP 429 → built-in backoff handles this, just wait

Parallel details:

• Each instance gets unique channels: orders_channel_instance_0, etc.
• Customer ID range (1-10000) is partitioned across instances
• Default batch: 10,000 orders per append_rows call
• Benchmark: 10K orders across 5 instances in ~12 seconds

Step 7: Verify Data

The parallel orchestrator waits 65 seconds after streaming before reconciliation. This prevents falsely detecting orphaned records.

SELECT COUNT(*) FROM <DATABASE>.<SCHEMA>.ORDERS;
SELECT COUNT(*) FROM <DATABASE>.<SCHEMA>.ORDER_ITEMS;

Expected ratio: ~3-4 order items per order.

Data may take 1-3 minutes to become queryable after run completes. If COUNT returns 0, wait and retry.

⚠️ STOP: Confirm data landed successfully.

Post-Run: Record Learnings

If the run had any deviations from what SKILL.md instructed (changed imports, added dependencies, modified configs, encountered undocumented errors), record them:

1. Read FEEDBACK_TEMPLATE.md for the entry format
2. Append an entry to LEARNINGS.md in the skill directory (not the user's project copy)
3. Set status to PENDING or APPLIED depending on whether you also patched the bundled files

If the run succeeded with no deviations, skip this — do not add a "no issues" entry.

Troubleshooting

SDK v1.2.0 API Reference

Import Path:

# CORRECT
from snowflake.ingest.streaming.streaming_ingest_client import StreamingIngestClient

# WRONG — does NOT exist:
# from snowpipe_streaming.client import StreamingIngestClient

Client Constructor — One Client Per Pipe:

orders_client = StreamingIngestClient(
    client_name="ORDERS_CLIENT_abc123",
    db_name="<DATABASE>",
    schema_name="<SCHEMA>",
    pipe_name="<ORDERS_TABLE>-STREAMING",
    properties={
        "account": "ORG-ACCOUNT",
        "user": "USERNAME",
        "private_key": open("key.p8").read(),  # FULL PEM with headers
        "role": "ACCOUNTADMIN",
        "url": "https://ORG-ACCOUNT.snowflakecomputing.com:443",
    }
)

Channel Operations:

channel, status = client.open_channel(channel_name="my_channel")

channel.append_rows(
    rows=[{"COL1": "val1"}],
    start_offset_token="orders_0",   # String tokens
    end_offset_token="orders_999",
)

channel.close(wait_for_flush=True)
client.close()

Private Key Format: Must be the full PEM string including -----BEGIN PRIVATE KEY----- headers. The bundled config_manager.py handles this automatically when private_key_file is set in the profile.

First-Time / Trial Account Setup

Skip this section if the user already has Cortex Code and a Snowflake connection configured.

Install Cortex Code

brew install Snowflake-Labs/cortex-code/cortex
# or: npm install -g @snowflake-labs/cortex-code

Set Up Snowflake Connection

If the streaming target is a different account from the user's current Cortex Code session, use key-pair auth (not externalbrowser). Browser auth silently routes to whatever account the user is logged into.

[streaming]
account = "orgname-accountname"
user = "YOUR_USERNAME"
authenticator = "SNOWFLAKE_JWT"
private_key_file = "/absolute/path/to/key.p8"
host = "orgname-accountname.snowflakecomputing.com"
database = "<DATABASE>"
schema = "<SCHEMA>"
warehouse = "<WAREHOUSE>"
role = "<ROLE>"

Finding your account identifier:

• Snowsight → click your name (bottom-left) → Account → copy the orgname-accountname value
• Uses hyphens not dots or underscores

Activate: /connection streaming

Architecture Notes

• Customer IDs are synthetic (1-10000) — no database lookup needed.
• One StreamingIngestClient per target table (not one for all tables).
• config_manager.py accepts both private_key (inline PEM) and private_key_file (path).
• Reconciliation connects via snowflake.connector with DER-encoded private key (separate from SDK).
• Parallel orchestrator waits 65s between streaming and reconciliation.
• Private key must be PKCS8 unencrypted (BEGIN PRIVATE KEY, not BEGIN RSA PRIVATE KEY).
• Role must be specified in profile — SDK does not default to any role.

Output

• Streaming project directory with configured venv, keys, and profiles
• Snowflake database/schema/tables created and owned by the specified role
• Synthetic e-commerce data (orders + order items) streamed into target tables
• Verified row counts confirming successful ingestion