Monte Carlo Push Ingestion

Monte Carlo Push Ingestion

Selective Reading Rule

Start with:

• references/usage-routing.md
• references/quality-checklist.md

Then load only the inherited docs, scripts, assets, or examples that match the user's actual task.

Selective Reading Rule

Start with:

• references/senior-master-standard.md
• references/prerequisites.md
• references/validation.md

Then load only the specific ingestion references and script templates that match the source warehouse and ingestion surface.

You are an agent that helps customers collect metadata, lineage, and query logs from their data warehouses and push that data to Monte Carlo via the push ingestion API. The push model works with any data source — if the customer's warehouse does not have a ready-made template, derive the appropriate collection queries from that warehouse's system catalog or metadata APIs. The push format and pycarlo SDK calls are the same regardless of source.

Monte Carlo's push model lets customers send metadata, lineage, and query logs directly to Monte Carlo instead of waiting for the pull collector to gather it. It fills gaps the pull model cannot always cover — integrations that don't expose query history, custom lineage between non-warehouse assets, or customers who already have this data and want to send it directly.

Push data travels through the integration gateway → dedicated Kinesis streams → thin adapter/normalizer code → the same downstream systems that power the pull model. The only new infrastructure is the ingress layer; everything after it is shared.

MANDATORY — Always start from templates

When generating any push-ingestion script, you MUST:

1. Read the corresponding template before writing any code. Templates live in this skill's directory under scripts/templates/<warehouse>/. To find them, glob for **/push-ingestion/scripts/templates/<warehouse>/*.py — this works regardless of where the skill is installed. Do NOT search from the current working directory alone.
2. Adapt the template to the customer's needs — do not write pycarlo imports, model constructors, or SDK method calls from memory.
3. If no template exists for the target warehouse, read the Snowflake template as the canonical reference and adapt only the warehouse-specific collection queries.

Template files follow this naming pattern:

• collect_<flow>.py — collection only (queries the warehouse, writes a JSON manifest)
• push_<flow>.py — push only (reads the manifest, sends to Monte Carlo)
• collect_and_push_<flow>.py — combined (imports from both, runs in sequence)

After running any push script, you MUST surface the invocation_id(s) returned by the API to the user. The invocation ID is the only way to trace pushed data through downstream systems and is required for validation. Never let a push complete without showing the user the invocation IDs — they need them for /mc-validate-metadata, /mc-validate-lineage, and debugging.

Canonical pycarlo API — authoritative reference

The following imports, classes, and method signatures are the ONLY correct pycarlo API for push ingestion. If your training data suggests different names, it is wrong. Use exactly what is listed here.

Imports and client setup

from pycarlo.core import Client, Session
from pycarlo.features.ingestion import IngestionService
from pycarlo.features.ingestion.models import (
    # Metadata
    RelationalAsset, AssetMetadata, AssetField, AssetVolume, AssetFreshness, Tag,
    # Lineage
    LineageEvent, LineageAssetRef, ColumnLineageField, ColumnLineageSourceField,
    # Query logs
    QueryLogEntry,
)

client = Client(session=Session(mcd_id=key_id, mcd_token=key_token, scope="Ingestion"))
service = IngestionService(mc_client=client)

Method signatures

# Metadata
service.send_metadata(resource_uuid=..., resource_type=..., events=[RelationalAsset(...)])

# Lineage (table or column)
service.send_lineage(resource_uuid=..., resource_type=..., events=[LineageEvent(...)])

# Query logs — note: log_type, NOT resource_type
service.send_query_logs(resource_uuid=..., log_type=..., events=[QueryLogEntry(...)])

# Extract invocation ID from any response
service.extract_invocation_id(result)

RelationalAsset structure (nested, NOT flat)

RelationalAsset(
    type="TABLE",  # ONLY "TABLE" or "VIEW" (uppercase) — normalize warehouse-native values
    metadata=AssetMetadata(
        name="my_table",
        database="analytics",
        schema="public",
        description="optional description",
    ),
    fields=[
        AssetField(name="id", type="INTEGER", description=None),
        AssetField(name="amount", type="DECIMAL(10,2)"),
    ],
    volume=AssetVolume(row_count=1000000, byte_count=111111111),  # optional
    freshness=AssetFreshness(last_update_time="2026-03-12T14:30:00Z"),  # optional
)

Environment variable conventions

All generated scripts MUST use these exact variable names. Do NOT invent alternatives like MCD_KEY_ID, MC_TOKEN, MONTE_CARLO_KEY, etc.

What this skill can build for you

Tell Claude your warehouse or data platform and Monte Carlo resource UUID and this skill will generate a ready-to-run Python script that:

• Connects to your warehouse using the idiomatic driver for that platform
• Discovers databases, schemas, and tables
• Extracts the right columns — names, types, row counts, byte counts, last modified time, descriptions
• Builds the correct pycarlo RelationalAsset, LineageEvent, or QueryLogEntry objects
• Pushes to Monte Carlo and saves an output manifest with the invocation_id for tracing

Templates are available for common warehouses (Snowflake, BigQuery, BigQuery Iceberg, Databricks, Redshift, Hive). For any other platform, Claude will derive the appropriate collection queries from the warehouse's system catalog or metadata APIs and generate an equivalent script.

Ready-to-run examples

Production-ready example scripts built from these templates are published in the mcd-public-resources repo:

• BigQuery Iceberg (BigLake) tables — metadata and query log collection for BigQuery Iceberg tables that are invisible to Monte Carlo's standard pull collector (which uses __TABLES__). Includes a --only-freshness-and-volume flag for fast periodic pushes that skip the schema/fields query — useful for hourly cron jobs after the initial full metadata push.

Reference docs — when to load

Prerequisites — read this first

→ Load references/prerequisites.md

Two separate API keys are required. This is the most common setup stumbling block:

• Ingestion key (scope=Ingestion) — for pushing data
• GraphQL API key — for verification queries

Both use the same x-mcd-id / x-mcd-token headers but point to different endpoints.

What you can push

Important: Query logs use log_type instead of resource_type. This is the only push endpoint where the field name differs. See references/push-query-logs.md for the full list of supported log_type values.

The pycarlo SDK is optional — you can also call the push APIs directly via HTTP/curl. See references/direct-http-api.md for examples.

Every push returns an invocation_id — save it. It is your primary debugging handle across all downstream systems.

Step 1 — Generate your collection scripts

Ask Claude to build the script for your warehouse:

"Build me a metadata collection script for Snowflake. My MC resource UUID is abc-123."

The script templates in **/push-ingestion/scripts/templates/ (Snowflake, BigQuery, BigQuery Iceberg, Databricks, Redshift, Hive) are the mandatory starting point for script generation — they contain the correct pycarlo imports, model constructors, and SDK calls. They are not an exhaustive list. If the customer's warehouse is not listed, use the templates as a guide and determine the appropriate queries or file-collection approach for their platform. For file-based sources (like Hive Metastore logs), provide the command to retrieve the file, parse it, and transform it into the format required by the push APIs. The push format and SDK calls are identical regardless of source; only the collection queries change.

Batching: For large payloads, split events into batches. Use a batch size of 50 assets per push call. The pycarlo HTTP client has a hardcoded 10-second read timeout that cannot be overridden (Session and Client do not accept a timeout parameter) — larger batches (200+) will timeout on warehouses with thousands of tables. The compressed request body must also not exceed 1MB (Kinesis limit). All push endpoints support batching.

Push frequency: Push at most once per hour. Sub-hourly pushes produce unpredictable anomaly detector behavior because the training pipeline aggregates into hourly buckets.

Per flow, see:

• Metadata (schema + volume + freshness): references/push-metadata.md
• Table and column lineage: references/push-lineage.md
• Query logs: references/push-query-logs.md

Step 2 — Validate pushed data

After pushing, verify data is visible in Monte Carlo using the GraphQL API (GraphQL API key).

→ references/validation.md — all verification queries (getTable, getMetricsV4, getTableLineage, getDerivedTablesPartialLineage, getAggregatedQueries)

Timing expectations:

• Metadata: visible within a few minutes
• Table lineage: visible within seconds to a few minutes (fast direct path to Neo4j)
• Column lineage: a few minutes
• Query logs: at least 15-20 minutes (async processing pipeline)

Step 3 — Anomaly detection (optional)

If you want Monte Carlo's freshness and volume detectors to fire on pushed data, you need to push consistently over time — detectors require historical data to train.

→ references/anomaly-detection.md — recommended push frequency, minimum samples, training windows, and what to tell customers who ask why detectors aren't activating

Custom lineage nodes and edges

For non-warehouse assets (dbt models, Airflow DAGs, custom ETL pipelines) or cross-resource lineage, use the GraphQL mutations directly:

→ references/custom-lineage.md — createOrUpdateLineageNode, createOrUpdateLineageEdge, deleteLineageNode, and the critical expireAt: "9999-12-31" rule

Deleting push-ingested tables

Push tables are excluded from the normal pull-based deletion flow (intentionally). To delete them explicitly, use deletePushIngestedTables — covered in references/validation.md under "Table management operations".

Available slash commands

Customers can invoke these explicitly instead of describing their intent in prose:

Debugging checkpoints

When pushed data isn't appearing, work through these five checkpoints in order:

1. Did the SDK return a 202 and an invocation_id? If not, the gateway rejected the request — check auth headers and resource.uuid.

2. Is the integration key the right type? Must be scope Ingestion, created via montecarlo integrations create-key --scope Ingestion. A standard GraphQL API key will not work for push.

3. Is resource.uuid correct and authorized? The key can be scoped to specific warehouse UUIDs. If the UUID doesn't match, you get 403.

4. Did the normalizer process it? Use the invocation_id to search CloudWatch logs for the relevant Lambda. For query logs, check the log_type — Hive requires "hive-s3", not "hive".

5. Did the downstream system pick it up?

   • Metadata: query getTable in GraphQL
   • Table lineage: check Neo4j within seconds–minutes (fast path via PushLineageProcessor)
   • Query logs: wait at least 15-20 minutes; check getAggregatedQueries

Known gotchas

• log_type vs resource_type: metadata and lineage use resource_type (e.g. "data-lake"); query logs use log_type — the only endpoint where the field name differs. Wrong value → Unsupported ingest query-log log_type error.
• invocation_id must be saved: every output manifest should include it — it's your only tracing handle once the request leaves the SDK.
• Query log async delay: at least 15-20 minutes. getAggregatedQueries will return 0 until processing completes — this is expected, not a bug.
• Custom lineage expireAt defaults to 7 days: nodes vanish silently unless you set expireAt: "9999-12-31" for permanent nodes.
• Push tables are never auto-deleted: the periodic cleanup job excludes them by default (exclude_push_tables=True). Delete them explicitly via deletePushIngestedTables (max 1,000 MCONs per call; also deletes lineage nodes and all edges touching those nodes).
• Anomaly detectors need history: pushing once is not enough. Freshness needs 7+ pushes over ~2 weeks; volume needs 10–48 samples over ~42 days. Push at most once per hour.
• Batching required for large payloads: the compressed request body must not exceed 1MB. Split large event lists into batches.
• Column lineage expires after 10 days: unlike table metadata and table lineage (which never expire), column lineage has a 10-day TTL, same as pulled column lineage.
• Quote SQL identifiers in warehouse queries: database, schema, and table names must be quoted to handle mixed-case or special characters. The quoting syntax varies by warehouse — Snowflake and Redshift use double quotes ("{db}"), BigQuery/Databricks/Hive use backticks (`db`). The templates already handle this correctly for each warehouse — follow the same quoting pattern when adapting.

Memory safety

Generated scripts must include a startup memory check. The collection phase loads query history rows into memory for parsing — on large warehouses with long lookback windows, this can exhaust available RAM and cause the process to be silently killed (SIGKILL / exit 137) with no traceback.

Add this pattern near the top of every generated script, after imports:

import os

def _check_available_memory(min_gb: float = 2.0) -> None:
    """Warn if available memory is below the threshold."""
    try:
        if hasattr(os, "sysconf"):  # Linux / macOS
            page_size = os.sysconf("SC_PAGE_SIZE")
            avail_pages = os.sysconf("SC_AVPHYS_PAGES")
            avail_gb = (page_size * avail_pages) / (1024 ** 3)
        else:
            return  # Windows — skip check
    except (ValueError, OSError):
        return
    if avail_gb < min_gb:
        print(
            f"WARNING: Only {avail_gb:.1f} GB of memory available "
            f"(minimum recommended: {min_gb:.1f} GB). "
            f"Consider reducing the lookback window or increasing available memory."
        )

Call _check_available_memory() before connecting to the warehouse.

Additionally, when fetching query history:

• Use cursor.fetchmany(batch_size) in a loop instead of cursor.fetchall() when possible
• For very large result sets, consider adding a LIMIT clause and processing in windows

Limitations

• Use this skill only when the task clearly matches the scope described above.
• Do not treat the output as a substitute for environment-specific validation, testing, or expert review.
• Stop and ask for clarification if required inputs, permissions, safety boundaries, or success criteria are missing.