How do I transfer data from MongoDB to BigQuery?

Pick a mechanism by what must land. Google Datastream replicates MongoDB into BigQuery continuously as JSON. The Dataflow MongoDB to BigQuery template runs a batch copy. Managed pipelines such as Fivetran and Airbyte sync on a schedule. A mapped sync lands typed columns. Then decide which fields become columns before the first load, not after.

Is there a MongoDB BigQuery connector?

Yes, several. Google offers two first-party routes: Datastream, which has supported MongoDB as a source since September 2025, and two Dataflow templates, one batch and one for change data. Fivetran, Airbyte, Estuary and Hevo sell managed connectors. They differ mostly in what lands: a JSON payload, top-level columns, or a schema inferred from a sample.

Can Datastream replicate MongoDB to BigQuery?

Yes. Datastream support for MongoDB as a source became generally available on 23 September 2025. It reads change streams, so it needs a replica set or sharded cluster, and Google documents support for MongoDB versions later than 5.0, including Atlas and self-managed servers. Documents arrive in BigQuery as Extended JSON, in canonical mode by default.

How does MongoDB to BigQuery CDC work?

A connector tails MongoDB change streams, the ordered feed of inserts, updates and deletes, and applies each event to BigQuery, either merged into the current row or appended as history. Change streams exist only on replica sets and sharded clusters. Google's Dataflow CDC template does not read them itself: you publish the events to Pub/Sub first.

What is the MongoDB to BigQuery Dataflow template?

It is a Google-provided batch pipeline that reads one collection and writes it to one BigQuery table. Its default, userOption NONE, writes three columns: id, the whole document as a JSON string in source_data, and a timestamp. FLATTEN promotes top-level fields to columns and JSON uses the native JSON type. Google lists it as Beta.

Can Airbyte sync MongoDB to BigQuery?

Yes. Airbyte reads MongoDB change streams and writes to BigQuery, where nested objects arrive in the native JSON type. By default it infers the schema from a sample of 10,000 documents, and its documentation says records with structural mismatches are written as NULL. Schemaless mode lands only _id and data instead, which keeps everything but types nothing.

Can Fivetran sync MongoDB to BigQuery?

Yes. Fivetran's MongoDB connector writes to BigQuery in packed mode by default, which keeps each document as JSON in a data column. Unpacked mode exists, but Fivetran states it unpacks only one layer of nested fields. Billing is by monthly active rows, plus a $5 base charge per connection that we read on Fivetran's pricing page.

Should MongoDB documents land in BigQuery as JSON or as columns?

Both, for most teams. Keep the raw document in a JSON column so nothing is ever lost, and project the fields your reports actually use into typed columns. BigQuery cannot partition or cluster a table on a JSON column, so a report that filters on a field buried in JSON scans more data than one that filters on a typed column.

How much does it cost to move MongoDB to BigQuery?

The pipeline bill is usually the smaller line. Datastream bills by the data it processes, Dataflow by worker time, Fivetran by monthly active rows plus a $5 base per connection, and Estuary by data moved plus connector instances. The larger cost is BigQuery compute spent unwrapping JSON on every query, which typed columns avoid.

MongoDB to BigQuery Connector, CDC and Dataflow

Which MongoDB to BigQuery connector should you use?

Choose by what has to land in BigQuery. If a JSON copy that your analysts unwrap in SQL is enough, Google Datastream is the first-party answer: it has replicated MongoDB change streams into BigQuery since September 2025. If you only need a one-off or nightly copy, the Dataflow batch template does it. If your reports need typed columns for nested fields, and most finance and product reports do, none of the defaults give you that. You either write and maintain the typed views yourself, or use a mapped sync that lands the declared paths as typed columns from the first load.

For the rest of the warehouse, the other sources and loaders are compared on BigQuery ETL tools, and if MongoDB is going to PostgreSQL instead, the same "what lands" question is answered on MongoDB to PostgreSQL migration tools.

Eight defaults, read from each vendor's own documentation

MongoDB never enforced a shape on your documents and BigQuery wants one for everything it can prune, cluster or aggregate cheaply. Every connector has to decide how to bridge that, and the setup wizard rarely tells you what it decided. The documentation does. The last column is what we would ship.

Tool and setting	What its documentation says	Why it matters to a buyer	What we would ship
Dataflow batch template, default	With userOption NONE, the table gets three columns: id, source_data and timestamp. Google states that source_data "contains the entire document as JSON", and the column type is STRING.	Every query parses a string. There is no type on anything, and no way to cluster on a field that lives inside it.	Use JSON mode at minimum, then project typed columns
Dataflow batch template, FLATTEN	Google's template docs describe FLATTEN as flattening "the documents to the single level", with top-level fields becoming table columns.	Only the top level becomes columns. customer.email and totals.grand_total stay nested, which is where most reporting fields live in a real document.	Declare nested paths explicitly
Dataflow CDC template	Google: "publish the change stream data to Pub/Sub. The pipeline reads the JSON records from Pub/Sub." The same page states it "doesn't support BigQuery change data capture".	The part that reads MongoDB is yours to build and run. And each change arrives as a new row, so a document updated forty times appears forty times until you deduplicate.	Merge on _id, keep history separately
Datastream, canonical mode	Canonical mode "ensures that every BSON type is explicitly labeled". A date is written as {"$date":{"$numberLong":"..."}}. Google's release notes date this default for new streams to 16 September 2026.	Nothing is lost, which is the point. But every date, long and decimal is a wrapper you unwrap in SQL, and streams created before 16 September wrote a different shape.	Typed view that handles both shapes
Datastream, strict mode	Google: "Strict mode doesn't support the Infinity, -Infinity, and NaN numeric values. Documents are discarded if they contain these values in strict mode."	A document with one NaN in one field does not arrive at all. The stream stays healthy, and the row count is simply lower than the collection count.	Scan for NaN before choosing strict
Airbyte, default mode	The MongoDB v2 source "enforces a schema" inferred from a sample, "Default is 10,000" documents, and records with structural mismatches "would be written as NULL".	A field the sample missed never becomes a column, and a value of the wrong shape becomes NULL. Both happen with every sync reporting success.	Inventory every field, not a sample
Airbyte, schemaless mode	Each record carries only "_id" and "data", and Airbyte notes that "no field will be omitted and no document will be rejected." In BigQuery, data lands as the native JSON type.	Complete and untyped. The same outcome as Datastream, with the modeling left to your analysts.	Fine as a raw layer, not as the model
Fivetran, default mode	Packed mode is "the default mode", writing data "without unpacking nested fields". In unpacked mode, "We only unpack one layer of nested fields and infer types."	The same payload landing as the others by default. Turning unpacking on gets you one layer, and anything deeper remains JSON.	Decide depth per path, not per connector

The row most teams will trip on this quarter is Datastream's. Canonical mode became the default for new streams on 16 September 2026, so an organization that created one stream in the summer and another this month now has two tables where the same date field sits at two different JSON paths. The SQL that unwraps both shapes, type by type, is in our guide to converting MongoDB Extended JSON into typed BigQuery columns. The Snowflake version of this route, where the first-party connector also lands a payload, is on MongoDB to Snowflake migration tools.

BSON types in BigQuery, and the wrong target a default picks

A MongoDB to BigQuery connector is mostly a type converter with Extended JSON in the middle. The second column is how the type travels. The last is the mistake we see most, and none of them raise an error.

BSON type	In Extended JSON	Right BigQuery target	Common wrong target
ObjectId	{"$oid": "..."} in every mode	STRING with the 24 hex characters	Leaving the $oid wrapper, so joins compare JSON objects
Date	Canonical: $date with $numberLong. Strict: $date as a number	TIMESTAMP, from UTC milliseconds	STRING, or DATETIME without a zone
Int64 (long)	{"$numberLong": "..."} as a string	INT64, cast from the string	FLOAT64 via a JSON number, past 2^53
Decimal128	{"$numberDecimal": "..."} as a string	NUMERIC, or BIGNUMERIC past 9 decimals	FLOAT64, which rounds currency
Double with NaN	Canonical: {"$numberDouble": "NaN"}	FLOAT64, which accepts NaN	Strict mode, which drops the document
Embedded document	A JSON object	Typed columns for used paths, JSON for the rest	FLATTEN, which stops at the top level
Array of subdocuments	A JSON array	ARRAY of STRUCT, or a child table with position	A JSON string nobody can UNNEST cheaply
Field with mixed types	Number in old documents, string in new	JSON kept, plus a view that casts deliberately	A typed column that turns mismatches into NULL
Missing field vs null	Absent key vs an explicit null	A nullable column, JSON kept to tell them apart	Collapsing both into NULL

One BigQuery rule changes the economics of the whole decision. The JSON type cannot be used to partition or cluster a table, because equality and comparison are not defined on it. A dashboard that filters orders by a status that lives inside JSON reads the whole table every time it refreshes. Promote the handful of fields you filter and join on into typed columns, and keep JSON for everything else. How BigQuery loads behave in general, including the Storage Write API, is covered in how to load data into BigQuery.

The six mechanisms behind "MongoDB to BigQuery"

Two of them are Google's own, which is unusual for a warehouse route and changes the buying question. You are rarely choosing whether a connector exists. You are choosing who writes the typed layer.

Mechanism	Owner	How it moves data	Who it is for	What to watch
Datastream	Google Cloud	Backfill plus continuous CDC from change streams, merged or append-only, as Extended JSON.	GCP shops that want a first-party, serverless stream.	Payload landing. Canonical and strict shapes differ.
Dataflow templates	Google Cloud	A batch template per collection, and a streaming template that reads change events you publish to Pub/Sub.	Teams already running Dataflow who want a copy, not a model.	Batch is Beta. CDC needs your own publisher.
Managed ELT	Fivetran, Airbyte, Estuary, Hevo	Reads change streams and writes to BigQuery on a schedule or continuously.	Teams that want zero pipeline operations.	Packed JSON or sampled schemas. Metered bills.
Open source CDC	You, on Debezium and Kafka Connect	Streams change events through Kafka into a BigQuery sink after a flattening transform.	Organizations already standardized on Kafka.	No license fee, real operational load.
Dump and load	You, on mongoexport and bq load	Exports collections to newline-delimited JSON, loads a staging table, reshapes with SQL.	One analysis on a few small collections.	Point in time only. No deletes, no updates.
Mapped sync (Adapters)	Adapters	You declare document paths once, we land typed columns and arrays as child tables, then sync incrementally.	Teams who want typed BigQuery tables on a flat bill.	Not a sub-second streaming platform.

Whichever mechanism you choose, the change feed underneath is the same MongoDB change stream. How it compares with log-based capture on relational databases, and what each source needs switched on, is on change data capture tools.

MongoDB to BigQuery connectors and CDC tools compared

Billing units rather than price tags, except where the vendor publishes the number itself: Fivetran's $5 base charge per connection and Estuary's $0.50 per GB were read from their own pricing pages in September 2026. Every other line depends on your volumes, so we state the unit and leave the arithmetic to a quote. Where a tool is wrong for a job, the last column says so, including for us.

Tool	Approach	Best for	Billing unit	What to watch
Google Datastream	Serverless CDC, merge or append-only	GCP teams wanting first-party replication	Data processed, plus BigQuery merge compute	Lands Extended JSON; you write the typed layer
Dataflow batch template	One collection to one table, NONE, FLATTEN or JSON	A one-off or nightly full copy	Dataflow worker time	Beta; FLATTEN is top level only
Dataflow CDC template	Pub/Sub JSON into BigQuery, streaming	Teams already publishing change events	Dataflow worker time plus Pub/Sub	You build the change stream reader
Fivetran	Managed connector, packed or one-layer unpacked	Teams who want no pipeline operations	Monthly active rows, plus $5 base per connection	Packed JSON by default
Airbyte	Change stream source, schema enforced or schemaless	Teams who want open source they can patch	Free self-hosted, credits on cloud	Mismatches written as NULL by default
Estuary	Streaming capture from change streams	Low latency with an easy setup	$0.50 per GB moved plus connector instances	Batch mode on standalone servers is not CDC
Hevo	Managed pipeline with automatic flattening	Smaller teams wanting a UI first	Plans by event volume	Ask how deep flattening goes
Debezium plus BigQuery sink	Change streams to Kafka to BigQuery	Organizations already running Kafka	Infrastructure and engineering time	Arrays and flattening are your decisions
mongoexport plus bq load	Export to JSON lines, load, reshape	One analysis on small collections	Free apart from BigQuery storage and queries	Relaxed JSON by default; no change capture
Adapters	Mapped sync of declared paths to typed columns	Typed reporting tables on a flat bill	Flat monthly plan from $49	Not built for sub-second latency

Datastream's own bill is only half of it. Google's destination documentation notes that BigQuery change data capture operations run on on-demand pricing by default, recommends a slot reservation for predictable cost, and exposes a max_staleness setting to trade freshness against merge cost. Ask for that line in any estimate. For how the managed vendors price themselves across every source, see the best data integration tools and Fivetran against Airbyte on cost.

Six numbers worth knowing before you sign anything

16 Sep 2026

The date Datastream made canonical Extended JSON the default for new MongoDB to BigQuery streams. Older streams may still write the strict shape.

Datastream release notes, read 26 Sep 2026

> 5.0

The MongoDB versions Datastream supports. It reads change streams, which exist only on replica sets and sharded clusters.

Datastream MongoDB source docs, read 26 Sep 2026

20 MB

The maximum event size Datastream writes into BigQuery. Extended JSON is wordier than BSON, so size the largest documents against it.

Datastream BigQuery destination docs, read 26 Sep 2026

3 columns

What the Dataflow batch template writes by default: id, source_data holding the whole document as a string, and timestamp.

Dataflow MongoDB to BigQuery template, read 26 Sep 2026

10,000

Airbyte's default sample for inferring a MongoDB schema. Values that do not fit the inferred shape are written as NULL.

Airbyte MongoDB v2 docs, read 24 Sep 2026

0

JSON columns you can partition or cluster a BigQuery table on. Filters on fields inside JSON cannot prune storage.

BigQuery JSON data docs, read 26 Sep 2026

Eight ways this pipeline goes wrong while every check stays green

Pipelines into BigQuery are built to keep going, and that is usually right. It also means the failures below never page anyone. We would check all eight before trusting a number from a new MongoDB feed.

The failure	What you see	What is actually happening
Two JSON shapes in one table	Old dashboards still work	A stream created before 16 September 2026 writes strict dates as a number, a new one writes canonical dates as a nested $numberLong string. SQL written for one shape returns NULL on the other, silently.
Documents discarded for NaN	Stream healthy, no errors	In strict mode Datastream discards any document holding NaN or Infinity. The collection count and the table count drift apart and nothing alerts.
The field the sample never saw	Sync green, counts match	Schema inference read 10,000 documents. A field that only appears later never became a column, even though every document was copied.
Mismatches written as NULL	Rows load, totals look low	A price stored as a string in old documents and a number in new ones becomes NULL wherever it does not fit the inferred type.
Forty rows for one order	Revenue is inflated	The Dataflow CDC template and append-only streams add a row per change. Without a view that keeps the latest row per _id, sums count every update.
FLATTEN stopped at the top	The table has columns	Only top-level fields became columns. The nested fields that reports need are still inside a JSON value or string.
Decimals through FLOAT64	Totals off by cents	Decimal128 extracted as a JSON number and read as FLOAT64 rounds. Extract the $numberDecimal string and cast it to NUMERIC instead.
The oplog gap after an outage	Stream resumes and carries on	Change streams resume from a position in the oplog. If the pipeline was down longer than the oplog window, those changes are gone unless you reconcile.

Five of the eight come from a shape that was guessed or inherited instead of declared. The guard is the same for all of them: a field inventory before the first load and a reconciliation of counts and NULLs after it. The one-time version of this problem, where you leave MongoDB entirely, is covered by data migration tools.

Six steps that decide whether the BigQuery tables are usable

Step 1

Inventory every field, not a sample

Run an aggregation over each full collection that lists every field path, how often it appears and which BSON types it holds. It replaces the sampling guess and tells you which fields deserve typed columns in BigQuery and which can stay in JSON.

Step 2

Confirm change streams are available

Every CDC route on this page needs a replica set or sharded cluster, and Datastream needs a MongoDB version later than 5.0. A standalone mongod rules out continuous replication until it is converted, so find this out in week one.

Step 3

Choose the landing shape per collection

Raw JSON for fast-changing collections, typed columns for the fields finance and product reports use, child tables for arrays of line items. Write the decision down. It is the part no connector makes for you.

Step 4

Pin the JSON format

If you use Datastream, set canonical or strict explicitly rather than inheriting a default that changed on 16 September 2026. Scan for NaN and Infinity first if you are tempted by strict mode.

Step 5

Backfill, then stream

Load history once, then follow the change stream. Decide early whether BigQuery should hold the current state of each document, the full change history, or both, because merge and append-only are set per stream.

Step 6

Reconcile counts and NULLs

Compare document counts per collection with distinct _id counts per table, and field presence in MongoDB with non-NULL counts in each column. This is the only check that catches discarded documents and mismatches written as NULL.

Why US teams pay to connect MongoDB to BigQuery

Product analytics on app data

The application runs on MongoDB Atlas and the analytics team works in BigQuery and Looker. They need orders, users and events joined, fresh within minutes, without querying production.

Finance reporting from a document store

Revenue, refunds and subscriptions live in nested documents. Finance needs them as typed NUMERIC columns that tie out to the ledger, not as strings parsed on every query.

Joining MongoDB with SaaS data

Customer records in MongoDB need to meet Salesforce, HubSpot and Stripe data already in BigQuery. The join keys have to be clean strings, not $oid wrappers.

Retiring a nightly export script

A cron job runs mongoexport and bq load every night and fails quietly once a month. The team wants change capture, retries and alerts instead of a script nobody owns.

Machine learning features

Data scientists want BigQuery ML or Vertex AI over behavioral data that lives in MongoDB. They need consistent types across years of documents written by different app versions.

Audit and retention

Compliance wants the full change history of certain collections kept for years. An append-only stream into BigQuery, with a current-state view on top, gives them both.

Five jobs where you should not pick us

A comparison page that never says the competition wins is an advert. These are the cases where something else on this page is the better buy.

You are a GCP shop and a JSON landing is genuinely enough for your analysts. Datastream is first-party, serverless and already in your console, and it bills by the data it processes. Use it and write the views.
You need a one-off copy of a few collections for a single analysis. The Dataflow batch template in JSON mode, or mongoexport and bq load, will do it for very little.
You already run Kafka and Debezium. Adding a BigQuery sink to a pipeline you operate today costs less than adding a vendor.
You need sub-second latency from a MongoDB write to a BigQuery row. Streaming platforms built for that will beat our incremental sync.
Your collections change shape weekly and nobody can say which fields matter yet. Land raw JSON first with any tool here and come back to typed columns once the reports settle.

Four questions to ask any vendor on this list

Question 01

What exactly lands in the BigQuery table?

Ask for the DDL of one real collection after a test sync. A three-column payload table, top-level columns only, a schema from a sample, or typed nested paths are all possible answers, and they lead to very different query bills.

Question 02

How are updates and deletes represented?

Merged into the current row, appended as a history row, or a soft-delete flag. Each is fine if you know which one you have. The expensive surprise is summing an append-only table as if it held current state.

Question 03

What happens to a value that does not match the column type?

An error you can see, a NULL you cannot, or the value kept in JSON beside the typed column. Two well-known tools on this page write NULL by default and say so in their own documentation.

Question 04

What do you need from our MongoDB deployment?

Change capture needs a replica set or sharded cluster, a supported version, network access from the vendor or from Google, and read roles. Any of these can add infrastructure work before a single document moves.

Related BigQuery and MongoDB guides

The typed-column guide for this route, the other MongoDB destinations, and the BigQuery pages for the rest of your sources.

Questions buyers ask about MongoDB to BigQuery

How do I transfer data from MongoDB to BigQuery?: Pick a mechanism by what must land. Google Datastream replicates MongoDB into BigQuery continuously as JSON. The Dataflow MongoDB to BigQuery template runs a batch copy. Managed pipelines such as Fivetran and Airbyte sync on a schedule. A mapped sync lands typed columns. Then decide which fields become columns before the first load, not after.
Is there a MongoDB BigQuery connector?: Yes, several. Google offers two first-party routes: Datastream, which has supported MongoDB as a source since September 2025, and two Dataflow templates, one batch and one for change data. Fivetran, Airbyte, Estuary and Hevo sell managed connectors. They differ mostly in what lands: a JSON payload, top-level columns, or a schema inferred from a sample.
Can Datastream replicate MongoDB to BigQuery?: Yes. Datastream support for MongoDB as a source became generally available on 23 September 2025. It reads change streams, so it needs a replica set or sharded cluster, and Google documents support for MongoDB versions later than 5.0, including Atlas and self-managed servers. Documents arrive in BigQuery as Extended JSON, in canonical mode by default.
How does MongoDB to BigQuery CDC work?: A connector tails MongoDB change streams, the ordered feed of inserts, updates and deletes, and applies each event to BigQuery, either merged into the current row or appended as history. Change streams exist only on replica sets and sharded clusters. Google's Dataflow CDC template does not read them itself: you publish the events to Pub/Sub first.
What is the MongoDB to BigQuery Dataflow template?: It is a Google-provided batch pipeline that reads one collection and writes it to one BigQuery table. Its default, userOption NONE, writes three columns: id, the whole document as a JSON string in source_data, and a timestamp. FLATTEN promotes top-level fields to columns and JSON uses the native JSON type. Google lists it as Beta.
Can Airbyte sync MongoDB to BigQuery?: Yes. Airbyte reads MongoDB change streams and writes to BigQuery, where nested objects arrive in the native JSON type. By default it infers the schema from a sample of 10,000 documents, and its documentation says records with structural mismatches are written as NULL. Schemaless mode lands only _id and data instead, which keeps everything but types nothing.
Can Fivetran sync MongoDB to BigQuery?: Yes. Fivetran's MongoDB connector writes to BigQuery in packed mode by default, which keeps each document as JSON in a data column. Unpacked mode exists, but Fivetran states it unpacks only one layer of nested fields. Billing is by monthly active rows, plus a $5 base charge per connection that we read on Fivetran's pricing page.
Should MongoDB documents land in BigQuery as JSON or as columns?: Both, for most teams. Keep the raw document in a JSON column so nothing is ever lost, and project the fields your reports actually use into typed columns. BigQuery cannot partition or cluster a table on a JSON column, so a report that filters on a field buried in JSON scans more data than one that filters on a typed column.
How much does it cost to move MongoDB to BigQuery?: The pipeline bill is usually the smaller line. Datastream bills by the data it processes, Dataflow by worker time, Fivetran by monthly active rows plus a $5 base per connection, and Estuary by data moved plus connector instances. The larger cost is BigQuery compute spent unwrapping JSON on every query, which typed columns avoid.

For the relational sources feeding the same warehouse, see replicating MySQL tables into BigQuery, and for the field-by-field SQL on this route read how to turn MongoDB Extended JSON into typed BigQuery columns.

MongoDB to BigQuery connector, CDC and Dataflow options compared, and what actually lands in the table