The executive question: should every dataset be copied into your warehouse?
Your finance team uses BigQuery. Product events sit in Amazon S3, operational extracts land in Azure Blob Storage, and customer profiles are managed in Salesforce. The conventional response is to build pipelines that copy everything into a central warehouse before anyone can join or analyze it. Sometimes that is the right answer. Sometimes it creates a costly second copy of data that is queried only occasionally.
BigQuery cross-cloud queries create another option: expose selected remote datasets through BigQuery and query them without first building a persistent, full-scale replication pipeline. That does not mean data never moves across networks. A remote read still transfers bytes for processing, may incur cloud-provider egress charges, and can create temporary or cached copies. The useful distinction is between reading remote data when needed and continuously maintaining a durable local replica.
The business decision is not “Can BigQuery connect?” It is “Which approach gives this workload acceptable freshness, predictable performance, governed access and the lowest sustainable operating cost?”
This guide provides an architecture comparison, a practical cost model, a security checklist and a four-week pilot that can be used before a production commitment.
Executive decision in one minute
- Favor direct cross-cloud access for exploratory analysis, occasional reconciliation, low-frequency joins and datasets whose ownership or structure makes replication undesirable.
- Favor managed replication or scheduled loading when the same remote data feeds frequent dashboards, heavy transformations, strict response-time commitments, historical snapshots or many downstream consumers.
- Consider BigQuery Omni when running supported processing near AWS or Azure data is materially better than moving raw data to a standard BigQuery region. Do not assume its feature set, regional support or billing is identical to newer cross-cloud connections.
- Consider a hybrid: query a large remote dataset selectively, then materialize a small, approved analytical table in BigQuery for repeated reporting.
- Require a measured pilot. Architecture diagrams cannot predict the real effects of file layout, cross-cloud egress, cache hit rates, source permissions and query concurrency.
The recommended default is workload-by-workload placement, not one policy for every source.
What BigQuery cross-cloud connections actually do
In August 2026, Google announced cross-cloud connections for querying AWS, Azure and Salesforce Data 360 from BigQuery regions in Preview. Google’s current cross-cloud connection documentation describes creating a connection in a standard BigQuery region, authorizing access to a remote source and querying remote tables or linked datasets. Confirm the current launch stage, eligible regions and applicable terms before production because product availability evolves.

For AWS, the documentation demonstrates external tables over S3 Parquet files and federated datasets backed by AWS Glue. For Azure, it shows external tables over Blob Storage. For Salesforce Data 360, it describes linked datasets created in a standard BigQuery location. These are distinct source-specific integration paths, not a universal SQL connection to every AWS database, Azure service or Salesforce object.
A cross-cloud connection in a standard BigQuery region reads remote data into the BigQuery execution environment. It can use ordinary BigQuery slot reservations rather than requiring separate Omni capacity. Google also documents optional cross-cloud caching of retrieved file blocks in the connection’s Google Cloud region. Caching is disabled by default and must be evaluated for freshness, data handling and cost.
There are three separate ideas that are often incorrectly grouped under “zero copy.”
1. Direct remote querying
A BigQuery external table or linked dataset references data held by another platform. You avoid maintaining a full persistent local copy solely to begin analysis. Query execution can still transfer data across cloud boundaries.

2. BigQuery Omni
Omni uses remote compute workers in supported AWS and Azure locations. It can reduce unnecessary raw-data movement for suitable workloads, but joins, materialization and other operations can still move selected results. Region compatibility, functionality and pricing must be evaluated separately.
3. Replication or loading
An ingestion process copies data into BigQuery tables on a schedule or incrementally. This increases storage and pipeline responsibilities but often gives repeatable performance, independent source availability, a stable transformation layer and control over historical states.
A fourth option, cross-cloud access to open table formats through Lakehouse catalogs, is relevant for Apache Iceberg ecosystems. It has its own governance and catalog considerations. Do not assume that an S3 Parquet external table, an Iceberg federated catalog and a Salesforce Data 360 linked dataset behave identically.
Start with the business workload, not the connector
An architect should inventory the questions the business actually asks. For each query or dashboard, record its owner, users, frequency, required freshness, maximum acceptable runtime, source size, selectivity, retention needs and sensitivity classification.
Consider three illustrative workloads. An analyst reconciles a remote monthly archive against a BigQuery finance table once per month. A customer-success dashboard refreshes every five minutes for hundreds of users. A risk team needs to review a narrow slice of last week’s remote events during an investigation. All three can reference the same external data but have very different economics.
The monthly reconciliation is a strong candidate for direct access. The high-frequency dashboard is likely better served by a curated local table or materialized subset. The investigation may benefit from remote querying with strict access controls and a temporary, expiring analytical result.
A useful first principle: optimize for the cost and reliability of the full decision workflow, not the elegance of avoiding a data pipeline.
AWS, Azure and Salesforce are three different integration problems
AWS: object storage, file layout and catalogs
For S3, first establish what is stored: Parquet, CSV, partitioned data, Iceberg tables or a Glue-cataloged collection. Columnar formats and selective predicates may reduce remote reads; unpartitioned CSV exports and SELECT * patterns can make remote querying expensive. Benchmark the actual files and SQL rather than assuming pushdown will behave as expected.
An S3 connection requires AWS IAM trust and access policies plus the BigQuery-side connection permissions. A sensible design grants read access to the intended prefixes, not the entire account. Verify the chosen AWS and BigQuery region pairing and the source account’s data-transfer charging rules.
Azure: Blob Storage and federated identity
For Azure Blob Storage, Google’s documented approach uses an Entra application, role assignments and federated credentials. Keep container access narrowly scoped, validate the remote path format, and test token or credential changes before rollout. Query cost will depend on the actual data scanned or fetched, not merely the number of SQL statements.
Salesforce: Data 360 is not the same as Sales Cloud
This distinction can prevent an expensive design error. The cross-cloud linked-dataset path is documented for Salesforce Data 360 (formerly Data Cloud), not as a generic direct-query interface to every Salesforce Sales Cloud or Service Cloud object. If your requirement is to analyze CRM Accounts, Opportunities or Cases from Sales Cloud, assess the supported Salesforce ingestion and transfer options separately.
Salesforce Data 360 sharing can reduce the need for a custom extraction process for eligible datasets, but you still need to test identity mapping, permitted data products, semantics, update timing and what happens when source access is revoked. A linked dataset is not a substitute for defining trusted business metrics.
Actiknow’s Salesforce integration and development services are relevant where CRM architecture and analytics requirements need to be assessed together; they do not imply that every Salesforce product supports the same cross-cloud query path.
A practical architecture comparison
Option A: Query remote data directly
Best fit: infrequent queries, selective investigation, ad hoc analysis, data that changes rapidly but is rarely queried, and sources whose owners prefer not to create a full replica.
Advantages: quicker access to some datasets, less duplicated storage, fewer scheduled ingestion jobs and potentially simpler source onboarding.
Trade-offs: remote-source dependence, network transfer charges, variable latency, possible source-side limits and more complex governance across two platforms.
Option B: Query with remote compute through Omni
Best fit: supported remote regions and workloads where processing close to the source reduces movement, subject to Omni’s documented capabilities.
Advantages: potential reduction in raw cross-cloud transfer for certain operations and familiar BigQuery SQL access.
Trade-offs: different region and capacity requirements, feature limitations relative to standard BigQuery execution and cross-region result-transfer considerations.
Option C: Replicate into BigQuery
Best fit: frequently reused data, heavy joins, predictable BI refreshes, consistent semantic modeling, long retention and repeated transformations.
Advantages: performance tuning against local storage, source outage isolation, governed history and reuse across teams.
Trade-offs: data movement and storage costs, connector operations, CDC or incremental logic, deletion handling and reconciliation obligations.
Option D: Hybrid selective materialization
Best fit: large remote sources with a small repeatedly used analytical subset.
Advantages: limited ingestion footprint and faster recurring reporting after materialization.
Trade-offs: requires a freshness policy, a transformation owner and explicit reconciliation of the subset against the remote source.

For many enterprises, Option D is the pragmatic middle ground. Query selectively, materialize what users repeatedly need, and keep the rest remote until usage justifies replication.
The real cost model: compute plus egress plus operations
Cross-cloud querying is not automatically cheaper because it avoids a scheduled ETL job. Evaluate total monthly cost:
Total cost = BigQuery query compute + remote cloud egress + remote request/read charges + optional cache-related charges + storage/materialization + engineering and support.
For replication, include initial and incremental transfer, ingestion compute, destination storage, transformations, monitoring and recovery. For direct access, include repeated remote reads, unpredictable concurrency, failed/retried queries and source-side availability.

An illustrative sensitivity calculation: suppose a query retrieves 60 GB of remote file blocks, runs 24 times per day and is executed 30 days per month. If every run retrieves the same amount, remote transfer exposure is 60 × 24 × 30 = 43,200 GB, or 43.2 TB using decimal units. This is not a price quote, a benchmark or a prediction of actual bytes fetched. Partition pruning, query planning, caching, file changes and repeated-block behavior can materially alter it.
Now compare a selective daily load of 60 GB once per day: 1.8 TB of inbound source data over the same period, plus local BigQuery storage and processing. Even if a direct query has no persistent replica, repeating it can cost more than loading a curated subset. Conversely, copying a huge archive for one small monthly reconciliation may be wasteful.
Build a sensitivity table with three cases: low, expected and peak. For each case estimate remote GB retrieved, queries per day, concurrency, cache hit rate, BigQuery slot usage, source egress rate, storage retained and engineer-hours required. Get actual cloud price inputs from the relevant billing accounts and regions.
Google’s BigQuery cross-cloud release notes record the original Preview launch and later changes, including November 2026 general availability of cross-cloud caching for Lakehouse access. Feature stage and cache stage are not necessarily the same; do not treat a caching announcement as proof that every connection type is generally available.
Latency: measure the entire user experience
A dashboard does not care whether a query is architecturally elegant. It cares how long a user waits and whether the numbers are current.
Measure at least five components: connection setup, remote read, network transfer, BigQuery processing and dashboard rendering. Record median and p95 end-to-end latency under both normal and peak concurrency. Averages can hide painful outliers.
Remote object-store layout matters. Large columnar files, partitioning on commonly filtered fields and stable schema can help; many tiny files, broad scans and poorly selective joins can hurt. But pushdown and caching depend on the specific table type and engine. Confirm with execution statistics rather than applying a generic optimization rule.
A remote dataset can be fresh while a cached block or materialized reporting table is older. Define freshness at the business measure level: for example, “orders visible within 30 minutes” or “finance closes at the previous day’s approved snapshot.” Then test the actual data-visibility path.
For critical BI, a sensible design is often to let a background job read or replicate remote data and publish a stable local mart. End users query the mart, not a live remote object store, during the board meeting.
Security and data residency: a query can still cross a boundary
Cross-cloud access does not eliminate the need for security review. The connection creates an identity bridge between providers, and the execution path may transfer source data into a different cloud or region.
Define exactly which principal can use the BigQuery connection, which remote objects it can read, who can query resulting tables, and who can create persistent outputs. Review row and column restrictions, sensitive attributes, audit logs, service-account rotation and revocation.
Data residency is especially important. A policy that prohibits storing a full replica outside a jurisdiction may also restrict transient processing or cache storage. Do not interpret “no permanent ingestion pipeline” as “no cross-border processing.” Document where the source data resides, where query execution occurs, whether caching is enabled and where derived results can be written.

A recommended control matrix has one row per dataset with: source owner, source region, BigQuery execution region, classification, connection identity, approved consumers, allowed purposes, cache policy, retention policy and audit owner.
Also test failure and offboarding. If a remote user, role or data product is revoked, does access stop as expected? If a BigQuery analyst loses permission, can they still see materialized outputs created earlier? Are temporary results and exports covered by retention rules?
Actiknow’s cloud application and multi-cloud integration services describe work across AWS, Azure and Google Cloud. That is relevant to the cross-provider security and operating model, not evidence of any specific unverified client implementation.
Reliability: remove the pipeline, not the operating discipline
A direct query is dependent on the source system, network path, permissions, object availability and BigQuery execution. When one layer is unavailable, a critical dashboard may fail even though the SQL and report are unchanged.
Define failure behavior before production. Should the report show the last approved snapshot? Retry with backoff? Fail closed with a visible freshness warning? Switch to a local fallback table? The right choice depends on whether a stale answer is safer than no answer.
Monitor source-read errors, authentication failures, bytes retrieved, query duration, cache behavior where applicable, freshness, quota failures and cost anomalies. Keep incident ownership clear across cloud teams.
Schema evolution also matters. A remote file adds a field or changes type; a source catalog changes a partition; a Salesforce data product is revised. Test whether queries fail loudly, silently omit fields or change metric meaning. An external connection does not remove the need for data contracts.
For business-critical numbers, reconcile remote results against source controls and any materialized reporting layer. Counts, financial totals, distinct keys, missing dates and join cardinality are more useful than “the query returned successfully.”
Worked implementation pattern: a selective AWS S3 pilot
The following SQL is an illustrative pattern adapted to Google’s documented S3 external-table approach. It is not a turnkey deployment script. The AWS IAM trust, connection, dataset location and source bucket permissions must already be configured and verified.
CREATE SCHEMA `demo_project.aws_raw`
OPTIONS (location = ‘us-east4’);
CREATE EXTERNAL TABLE `demo_project.aws_raw.sales_parquet`
WITH CONNECTION `us-east4.aws_readonly`
OPTIONS (
format = ‘PARQUET’,
uris = [‘s3://example-analytics/sales/year=2026/*’]
);
SELECT order_date, SUM(net_amount) AS revenue
FROM `demo_project.aws_raw.sales_parquet`
WHERE order_date >= DATE ‘2026-10-01’
AND order_date < DATE ‘2026-11-01’
GROUP BY order_date;
The identifiers and bucket are placeholders. The SQL is intended to show where the connection and remote path are referenced; it does not demonstrate a production-ready schema, verified partition pruning or a particular performance result.
Run the same business query against a separately loaded BigQuery table. Compare correctness first, then median and p95 latency, bytes read remotely, total cost and failure behavior. Repeat with different predicates and concurrent users. Do not accept one warm-cache execution as the performance baseline.
If the workload repeatedly reads the same narrow subset, test materializing only the approved columns and date range into a partitioned local table. Measure the refresh job and reporting queries together. This may be the most economical production design.
A four-week pilot that produces an architecture decision
Week 1: inventory and governance
Select two or three representative workloads: one exploratory, one scheduled BI, one high-volume or sensitive. Confirm source owners, data classification, business metrics, region constraints, refresh SLAs and the currently supported connection types. Capture baseline ETL or replication costs if they exist.
Deliverable: a signed workload inventory and a shortlist of feasible architectures.
Week 2: build the minimum viable connection
Create least-privilege connections for one AWS or Azure source and, if appropriate, one Salesforce Data 360 data product. Run read-only queries with explicit predicates. Validate identity mapping, query-region selection, error handling and auditability. Keep caching off initially so baseline remote-read behavior is visible, then test it as a separate scenario where supported.
Deliverable: reproducible configuration and access tests, with no unapproved production exposure.
Week 3: benchmark and reconcile
Run representative SQL under realistic frequency and concurrency. Record p50 and p95 latency, remote bytes, slot consumption, cloud egress, freshness and failure rates. Compare a loaded local copy and a selective materialized subset. Reconcile important metrics against source totals and document every difference.
Deliverable: an evidence-backed comparison, not a slide with hypothetical savings.
Week 4: decide and operationalize
Approve direct access, replication, Omni or hybrid separately for each workload. Document access owners, fallback behavior, alert thresholds, data contracts, billing controls, refresh windows and the conditions that trigger architectural reassessment.
Deliverable: a production decision record with measurable acceptance criteria and an exit plan.
A scorecard executives can use
Score each candidate from 1 (poor) to 5 (strong) using evidence from the pilot, not vendor claims.
- Freshness: Does the actual result meet the business SLA?
- Performance: Does p95 response time meet the user’s expectation at peak load?
- Total cost: Are compute, egress, storage and support predictable under growth?
- Governance: Are access, residency, classification and audit requirements satisfied?
- Reliability: Can the business tolerate a source or network outage?
- Reuse: Will many teams repeatedly consume the same curated data?
- Change tolerance: Can schema and permission changes be detected and handled safely?
- Reversibility: Can the workload move to a local replica without rebuilding the business definition?
Weight the dimensions according to the use case. A finance close may prioritize correctness and recoverability; exploratory research may prioritize access speed and flexibility. The scorecard should explain a decision, not manufacture a false numerical certainty.
Common architectural mistakes
“No ETL” means “no data movement”
It does not. Remote data can cross provider boundaries during query execution and caching. Distinguish persistent replication from transient reads.
“We connected Salesforce” means “all CRM objects are available”
A Data 360 linked dataset is not the same as a Sales Cloud API or the BigQuery Data Transfer Service connector. Check the actual source product and supported object scope.
“One successful query proves the design”
It does not test concurrency, data changes, cold-cache behavior, cost spikes, outages or month-end reporting.
“A single global architecture will simplify governance”
Different sources have different residency, ownership, update and cost characteristics. A documented decision policy is often simpler than forcing everything through one technical mechanism.
“Remote access removes data engineering”
It may remove some ingestion work, but not semantic modeling, reconciliation, data-quality testing, cost governance or incident response.
For executive reporting, Actiknow’s business intelligence and data integration approach emphasizes requirements, solution architecture, modeling, integrations and ongoing refresh management. Those responsibilities remain whether data is copied or queried remotely.
Frequently asked questions
1. Can BigQuery query AWS S3 without copying the entire bucket?
Yes, supported cross-cloud connections can expose S3 data through external tables. Queries still retrieve data over the network, and egress or caching costs may apply. A full persistent replica is not required merely to run the query.
2. Can BigQuery query Azure Blob Storage directly?
Google documents Azure cross-cloud connections and external tables over Blob Storage. Availability, permissions, region pairing and source formats should be confirmed for the specific workload.
3. Does this work with Salesforce Sales Cloud?
Do not assume so. The cross-cloud linked-dataset documentation refers to Salesforce Data 360. Sales Cloud data can require a different transfer, API or integration architecture.
4. Is BigQuery Omni the same as a cross-cloud connection?
No. Omni uses remote compute in supported locations. Newer cross-cloud connections can use standard BigQuery regional compute and reservations. Feature, region and pricing differences matter.
5. Is querying remote data always cheaper than ETL?
No. Repeated remote reads and egress can outweigh the cost of a curated local copy. Compare monthly cost under realistic query frequency and concurrency.
6. Will the data always be real time?
Not necessarily. Source update frequency, file publication, catalog metadata, caching and materialization determine effective freshness. Test a defined business freshness SLA.
7. Can we join AWS and BigQuery data?
Supported external data can participate in BigQuery queries and joins, subject to product and region restrictions. Test actual execution plans and transferred bytes before promising production performance.
8. Does cross-cloud access satisfy data residency requirements?
Not automatically. Remote bytes may be processed or cached in a Google Cloud region. Compliance review must consider execution location, derived outputs and contractual restrictions.
9. Should we enable cross-cloud caching?
Only after testing repeated-query benefit, cache behavior, freshness requirements and the security implications of storing remote file blocks locally. The feature is optional and disabled by default in the documented connection setup.
10. What should we monitor after launch?
Query latency, remote bytes, source errors, authentication failures, freshness, concurrency, egress charges, cache metrics where available and source-to-report reconciliation.
11. When should we switch from direct access to replication?
Reassess when query frequency, repeated scan volume, p95 latency, consumer count, transformation complexity or availability requirements make a local serving layer more reliable or economical.
12. Can a hybrid architecture become the long-term design?
Yes. Keep low-frequency and archival data remote, materialize frequently reused subsets, and publish governed business metrics from stable analytical models. Hybrid is a deliberate architecture, not an incomplete migration.
The decision that matters
Cross-cloud querying is valuable because it lets an enterprise choose when data should move, rather than treating copying as the prerequisite for every analytical question. Its value disappears if remote queries create unpredictable bills, violate residency rules or make executive dashboards depend on an unavailable source.
Start with the workload. Benchmark remote reads against replication and selective materialization. Include the cost of egress, governance, freshness and incident handling. Then choose the smallest architecture that meets the business requirement.
If your organization is deciding how to combine AWS, Azure, Salesforce and BigQuery reporting, Actiknow can help assess data sources, integration choices and BI delivery requirements. Discuss your data architecture with Actiknow to scope a focused technical and commercial evaluation.

