Sync HubSpot CRM data to a data warehouse (BigQuery, Snowflake, or Postgres) for
analytics and reporting. Covers initial backfill of millions of records under the
500K/day rate limit, incremental CDC polling via hs_lastmodifieddate, schema-drift
detection with ALTER TABLE generation, association sync for contacts/deals/companies,
and idempotent upsert patterns that prevent duplicate rows on retry. Use when building
a HubSpot → warehouse pipeline, resyncing after a schema change, debugging duplicate
rows or missing CDC updates, or recovering from a rate-limit burnout mid-backfill.
Trigger with "hubspot warehouse sync", "hubspot bigquery", "hubspot snowflake",
"hubspot postgres etl", "hubspot backfill", "hubspot cdc", "hubspot schema drift",
"hubspot duplicate rows", "hubspot to data warehouse".
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Sync HubSpot CRM data to a data warehouse (BigQuery, Snowflake, or Postgres) for
analytics and reporting. Covers initial backfill of millions of records under the
500K/day rate limit, incremental CDC polling via hs_lastmodifieddate, schema-drift
detection with ALTER TABLE generation, association sync for contacts/deals/companies,
and idempotent upsert patterns that prevent duplicate rows on retry. Use when building
a HubSpot → warehouse pipeline, resyncing after a schema change, debugging duplicate
rows or missing CDC updates, or recovering from a rate-limit burnout mid-backfill.
Trigger with "hubspot warehouse sync", "hubspot bigquery", "hubspot snowflake",
"hubspot postgres etl", "hubspot backfill", "hubspot cdc", "hubspot schema drift",
"hubspot duplicate rows", "hubspot to data warehouse".
Move HubSpot CRM data to BigQuery, Snowflake, or Postgres in a way that survives production — not just the demo. This is not a connector walkthrough. It is the extraction and load code your pipeline runs when a 2M-contact backfill burns through the 500K daily call quota at noon, when CDC misses three days of deal updates because association changes do not update hs_lastmodifieddate, when a portal admin adds a custom property and your warehouse table schema silently drifts, and when a network timeout at record 45,000 causes your retry to insert 100 duplicate rows.
The six production failures this skill prevents:
Backfill exhausting the daily rate limit before completing — 2M contacts at 100 records/call is 20,000 calls. At 100 calls/10s the math says 33 minutes, but that burns your entire 500K daily quota before noon and takes every other integration down with it. Token bucket rate limiting with a configurable daily ceiling is non-optional.
CDC missing association changes — HubSpot's hs_lastmodifieddate is updated when any property on the contact record changes, but not when an association is created or deleted. A contact-to-deal link added by a sales rep is invisible to a lastmodifieddate-based incremental poll. Association CDC requires a separate poll strategy.
Schema drift causing silent extraction failures — when a portal admin adds or removes a custom property, the warehouse table schema and the extraction property list become misaligned. New properties are dropped on the floor. Removed properties cause KeyError on row construction. Neither failure raises an alarm without explicit schema validation.
Duplicate rows on batch retry — a network failure mid-batch causes the batch to be re-sent. Without a proper upsert key the warehouse gets duplicate rows that are invisible until an analyst notices double-counted revenue. The correct upsert key for contacts is id (HubSpot object ID), not a composite of name/email.
Large payload failures on associated object inline pulls — contacts with thousands of engagement records cause payload sizes that exceed HTTP response limits when associations are pulled inline. Associations must be fetched in a separate batch read pass.
Timezone inconsistency in aggregations — HubSpot stores all timestamps as Unix milliseconds in UTC. If the warehouse session timezone is set to a local timezone, DATE(created_at) aggregations produce different daily totals depending on where the analyst runs the query.
Prerequisites
Python 3.10+
HubSpot private app token with scopes: crm.objects.contacts.read, , , ,
crm.objects.companies.read
crm.objects.deals.read
crm.associations.read
crm.schemas.contacts.read
Target warehouse credentials:
BigQuery: service account JSON with bigquery.dataEditor role
Snowflake: user/password or key-pair auth, warehouse + database + schema
Postgres: connection string with write access to the target schema
Python packages: requests, pandas, pyarrow, google-cloud-bigquery (BigQuery) or snowflake-connector-python (Snowflake) or psycopg2-binary (Postgres)
A secret store the runtime can read (env vars, GCP Secret Manager, AWS Secrets Manager)
Instructions
Build in this order. Each section neutralizes one production failure mode.
Naive extraction loops call the API as fast as possible. At 100 calls/10s with 20K calls needed, you burn 200K of your 500K daily quota in 33 minutes — and that is before any CDC polling or webhook processing. A production backfill must budget its calls over the full day so other integrations still have headroom.
The token bucket strategy: set a daily_budget ceiling below the account limit (e.g., 400K of 500K), compute how many calls per second that allows over 24 hours, and enforce a minimum interval between calls. Burn-rate is checked against live rate-limit headers on every response.
import time
import threading
from dataclasses import dataclass, field
@dataclassclassTokenBucket:
"""
Token bucket rate limiter for HubSpot API calls.
daily_budget: max calls to spend per 24h window (stay below 500K limit)
burst_limit: max calls per 10s window (100 for most tiers)
"""
daily_budget: int = 400_000
burst_limit: int = 95# leave 5 calls headroom per 10s window
_lock: threading.Lock = field(default_factory=threading.Lock)
_calls_today: int = 0
_window_calls: int = 0
_window_start: float = field(default_factory=time.monotonic)
_day_start: float = field(default_factory=time.monotonic)
defacquire(self) -> None:
withself._lock:
now = time.monotonic()
# Reset daily counterif now - self._day_start >= 86_400:
self._calls_today = 0self._day_start = now
# Reset 10s window counterif now - self._window_start >= 10.0:
self._window_calls = 0self._window_start = now
# Hard stop if daily budget exhausted — do not burn other integrationsifself._calls_today >= self.daily_budget:
seconds_left = 86_400 - (now - self._day_start)
raise DailyBudgetExhausted(
f"Daily call budget of {self.daily_budget:,} reached. "f"Resuming in {seconds_left/3600:.1f}h."
)
# Burst window throttle — sleep until window resets if fullifself._window_calls >= self.burst_limit:
sleep_s = 10.0 - (now - self._window_start) + 0.1
time.sleep(max(0, sleep_s))
self._window_calls = 0self._window_start = time.monotonic()
self._calls_today += 1self._window_calls += 1defupdate_from_headers(self, headers: dict) -> None:
"""Adjust pacing from live rate-limit headers on each response."""
remaining = int(headers.get("X-HubSpot-RateLimit-Daily-Remaining", self.daily_budget))
if remaining < 50_000:
# Emergency throttle: burn rate is too high, cut burst limit in halfwithself._lock:
self.burst_limit = max(10, self.burst_limit // 2)
classDailyBudgetExhausted(Exception):
pass
RATE_LIMITER = TokenBucket()
2. Full backfill with cursor pagination (contacts search API)
The search API (POST /crm/v3/objects/contacts/search) supports cursor pagination via after. It returns a maximum of 100 records per page and up to 10,000 records total per search. For tables larger than 10K records, use the lastmodifieddate range-slicing strategy: paginate within 30-day windows, sliding forward from the earliest hs_createdate in the portal.
import requests
import json
BASE_URL = "https://api.hubapi.com"deffetch_contacts_page(
token: str,
after: str | None,
properties: list[str],
rate_limiter: TokenBucket,
) -> tuple[list[dict], str | None]:
"""Fetch one page of contacts. Returns (records, next_cursor)."""
RATE_LIMITER.acquire()
body = {
"limit": 100,
"properties": properties,
"sorts": [{"propertyName": "hs_lastmodifieddate", "direction": "ASCENDING"}],
}
if after:
body["after"] = after
resp = requests.post(
f"{BASE_URL}/crm/v3/objects/contacts/search",
headers={"Authorization": f"Bearer {token}", "Content-Type": "application/json"},
json=body,
timeout=30,
)
rate_limiter.update_from_headers(dict(resp.headers))
if resp.status_code == 429:
retry_after = int(resp.headers.get("Retry-After", 10))
time.sleep(retry_after)
return fetch_contacts_page(token, after, properties, rate_limiter) # one retry
resp.raise_for_status()
data = resp.json()
results = data.get("results", [])
next_cursor = data.get("paging", {}).get("next", {}).get("after")
return results, next_cursor
defbackfill_contacts(
token: str,
properties: list[str],
rate_limiter: TokenBucket,
checkpoint_file: str = "/tmp/hubspot_backfill_checkpoint.json",
) -> int:
"""
Full backfill with checkpoint. Resume-safe: reads cursor from checkpoint_file
so a mid-run failure restarts from the last completed page, not from zero.
Returns total records written.
"""# Load checkpointtry:
withopen(checkpoint_file) as f:
checkpoint = json.load(f)
after = checkpoint.get("after")
total = checkpoint.get("total", 0)
print(f"Resuming backfill from cursor {after}, {total:,} records already written")
except FileNotFoundError:
after = None
total = 0whileTrue:
records, next_cursor = fetch_contacts_page(token, after, properties, rate_limiter)
ifnot records:
breakyield records # caller handles warehouse write
total += len(records)
after = next_cursor
# Write checkpoint after every page so a restart costs at most 100 recordswithopen(checkpoint_file, "w") as f:
json.dump({"after": after, "total": total}, f)
ifnot next_cursor:
breakprint(f"Backfill complete: {total:,} contacts")
return total
3. Incremental CDC poll (neutralizes missed updates — with association gap)
The standard CDC pattern polls on hs_lastmodifieddate > last_run. This catches property changes but silently misses association changes (linking a contact to a deal or removing that link does not update hs_lastmodifieddate). The fix is a two-pass incremental: a property poll and a separate association poll on a shorter interval.
import datetime
defincremental_contacts_since(
token: str,
since_ms: int,
properties: list[str],
rate_limiter: TokenBucket,
) -> list[dict]:
"""
CDC property poll: return all contacts modified after since_ms (Unix ms UTC).
Does NOT cover association changes — run poll_association_changes() separately.
"""
all_records = []
after = NonewhileTrue:
RATE_LIMITER.acquire()
body = {
"limit": 100,
"properties": properties,
"filterGroups": [{
"filters": [{
"propertyName": "hs_lastmodifieddate",
"operator": "GT",
"value": str(since_ms),
}]
}],
"sorts": [{"propertyName": "hs_lastmodifieddate", "direction": "ASCENDING"}],
}
if after:
body["after"] = after
resp = requests.post(
f"{BASE_URL}/crm/v3/objects/contacts/search",
headers={"Authorization": f"Bearer {token}", "Content-Type": "application/json"},
json=body,
timeout=30,
)
rate_limiter.update_from_headers(dict(resp.headers))
resp.raise_for_status()
data = resp.json()
all_records.extend(data.get("results", []))
next_cursor = data.get("paging", {}).get("next", {}).get("after")
ifnot next_cursor:
break
after = next_cursor
return all_records
defpoll_association_changes(
token: str,
contact_ids: list[str],
rate_limiter: TokenBucket,
) -> dict[str, list[str]]:
"""
Fetch current contact → deal associations for a list of contact IDs.
Use this to detect additions and deletions that hs_lastmodifieddate misses.
Strategy: compare fetched associations against warehouse snapshot.
Differences = changes that must be written.
Returns {contact_id: [deal_id, ...]}
"""# Batch read associations: up to 100 contacts per call (v4 associations API)
result = {}
for chunk_start inrange(0, len(contact_ids), 100):
chunk = contact_ids[chunk_start:chunk_start + 100]
RATE_LIMITER.acquire()
resp = requests.post(
f"{BASE_URL}/crm/v4/associations/contacts/deals/batch/read",
headers={"Authorization": f"Bearer {token}", "Content-Type": "application/json"},
json={"inputs": [{"id": cid} for cid in chunk]},
timeout=30,
)
rate_limiter.update_from_headers(dict(resp.headers))
resp.raise_for_status()
for item in resp.json().get("results", []):
from_id = str(item["from"]["id"])
to_ids = [str(a["toObjectId"]) for a in item.get("to", [])]
result[from_id] = to_ids
return result
4. Schema drift detection and ALTER TABLE generation
When a portal admin adds or removes a custom property, your warehouse table schema diverges from the API response silently. The correct mitigation is a pre-run schema check: enumerate all HubSpot properties via the properties API, diff against the warehouse column list, and emit ALTER TABLE statements for any new columns. Removed properties become nullable and are not dropped (dropping columns in production is a separate review).
deffetch_hubspot_property_schema(token: str) -> dict[str, str]:
"""
Returns {property_name: warehouse_type} for all contact properties.
Maps HubSpot field types to warehouse-appropriate column types.
"""
resp = requests.get(
f"{BASE_URL}/crm/v3/properties/contacts",
headers={"Authorization": f"Bearer {token}"},
timeout=30,
)
resp.raise_for_status()
TYPE_MAP = {
"string": "TEXT",
"number": "FLOAT64",
"date": "DATE",
"datetime": "TIMESTAMP",
"bool": "BOOLEAN",
"enumeration": "TEXT",
"phone_number":"TEXT",
"json": "TEXT", # store complex types as serialized JSON
}
schema = {}
for prop in resp.json().get("results", []):
hs_type = prop.get("type", "string")
schema[prop["name"]] = TYPE_MAP.get(hs_type, "TEXT")
return schema
defdetect_schema_drift(
hubspot_schema: dict[str, str],
warehouse_columns: set[str],
) -> tuple[dict[str, str], set[str]]:
"""
Returns:
added: {column_name: type} — in HubSpot but not warehouse
removed: {column_name} — in warehouse but not HubSpot
"""
hs_columns = set(hubspot_schema.keys())
added = {k: v for k, v in hubspot_schema.items() if k notin warehouse_columns}
removed = warehouse_columns - hs_columns - {"_synced_at", "_sync_run_id"} # exclude meta colsreturn added, removed
defgenerate_alter_statements(table: str, added: dict[str, str]) -> list[str]:
"""Generate ALTER TABLE ADD COLUMN statements for new HubSpot properties."""
stmts = []
for col, dtype in added.items():
safe_col = col.replace("-", "_").lower()
stmts.append(f"ALTER TABLE {table} ADD COLUMN IF NOT EXISTS {safe_col}{dtype};")
return stmts
5. Warehouse upsert patterns (neutralizes duplicate rows on retry)
Each warehouse requires a different upsert idiom. The upsert key is always id — HubSpot's immutable object ID. Never composite on mutable fields like email or name; those can change and cause phantom duplicates.
All three patterns use a staging-table approach: load new rows into a temp table, then merge into production — the most portable pattern across warehouse engines. Normalize records first with every timestamp parsed as Unix-millisecond UTC (never the session's local timezone) and missing properties filled with None, then merge on id:
6. Large association payloads (separate fetch pass)
Never pull associations inline in the contacts search response. A contact with 500 engagement records produces a response payload measured in megabytes that times out or exceeds HTTP limits. Fetch associations in a second pass using the batch associations endpoint, keyed on the contact IDs from step 2/3.
The implementation is in poll_association_changes() above. The rule is: contacts page first, IDs collected, associations fetched as a second batch read. Write associations to a separate hubspot_contact_deal_associations table — not as columns on the contacts table.
API_REFERENCE.md — search API cursor shapes, batch read request/response, CDC fields, rate-limit headers, association batch format
implementation-guide.md — complete backfill script with token bucket, full CDC poll loop, schema drift handler, ALTER TABLE generator, all three warehouse upsert implementations
