Actiknow
Data Engineering

BigQuery Lakehouse Caching: Can Repeated Cross-Cloud Queries Reduce Egress Costs?

Learn when BigQuery cross-cloud caching reduces repeat data transfers, how to measure cache hits, and when federation or replication makes more sense.

Cloud architecture and finops teams reviewing bigquery lakehouse caching for cross cloud analytics

The promise of querying data across clouds without copying it is attractive: fewer pipelines, faster access to distributed datasets and less pressure to consolidate every system into one warehouse. But “no migration” does not mean “no movement.” When BigQuery reads files stored in another cloud, bytes can still cross provider boundaries, creating network charges and latency.

BigQuery Lakehouse caching addresses a specific part of that problem. It keeps previously retrieved remote file blocks in the Google Cloud query region so later queries can reuse them rather than fetch the same blocks again. For analytics leaders, the decision is not simply whether to enable a cache. It is whether repeated reads, source-change patterns, security requirements and total workload costs make federation economically preferable to copying selected data.

What changed in 2026?

Google Cloud announced Lakehouse caching in Preview on August 31, 2026. On November 2, 2026, Google Cloud announced general availability for cross-cloud data access caching. As of this article, caching is disabled by default and must be enabled on a supported Lakehouse federated catalog or BigQuery cross-cloud connection. That distinction matters: an architecture designed around the original Preview announcement should be rechecked against the current generally available configuration and regional support.

The documented capability is a read-through block cache, not a new permanent enterprise data warehouse. When a query requests remote file blocks, the service can store retrieved blocks in the region where the query runs. Subsequent queries that need those same blocks may read them locally. Different SQL statements can benefit if they access overlapping file blocks; identical SQL text is not required.

It is equally important to distinguish this mechanism from BigQuery’s query-result cache. Result caching can avoid re-executing a qualifying query, while Lakehouse caching reduces remote object-storage reads during queries that still execute. Neither should be assumed to replace a governed reporting layer.

Cloud architect reviewing remote file block caching between cloud storage and bigquery query regions

Why egress becomes an executive issue

Consider an organization with operational data in Amazon S3 and its executive dashboards in BigQuery. The same monthly files might be scanned repeatedly by finance, operations and planning teams. Each remote read can contribute to cross-cloud transfer cost, even when the underlying data barely changes. Network latency can also make otherwise efficient SQL feel slow.

Caching is potentially valuable when several workloads revisit the same remote blocks. Its value is lower when each query touches entirely new files, when source files are frequently replaced, or when data is rarely queried more than once. A cache cannot make a one-time full historical scan free.

The right cost question is: How much paid remote transfer can we avoid, and what does that save after accounting for query compute, connection and network arrangements, operational overhead and any materialized alternatives?

A practical cost model

Start with measured remote bytes, not the size of the entire source dataset. Define:

  • R = remote bytes that would be fetched over the evaluation period without caching.
  • H = fraction of those bytes served from the local block cache when enabled.
  • P = effective remote egress price per byte, based on the actual provider, region, contract and network route.
  • C = incremental operational or infrastructure cost attributable to the chosen architecture.

A first-order estimate is avoided remote egress cost = R × H × P. Net benefit is avoided egress minus incremental costs, with any compute or latency savings evaluated separately. This is a planning model, not a billing guarantee: providers have tiered pricing, discounts, minimums and network-specific rules.

Illustrative scenario, not a customer result: a workload would otherwise retrieve 12 TB of remote file blocks in a month. If measured cache behavior avoids 60% of those reads, approximately 7.2 TB of remote reads may be avoided. Multiply that by the applicable effective transfer rate to estimate savings. Do not assign a universal dollar value until finance confirms the actual rate.

A crucial measurement warning: a 60% cache-byte share is not automatically a 60% reduction in the entire cloud bill. Query compute, storage, other transfers and fixed commitments remain.

FinOps professionals evaluating cross-cloud egress savings against BigQuery compute and infrastructure costs

How to measure actual cache use

Google Cloud documents job-level object-storage statistics. Inspect BigQuery job details or retrieve a completed job with the bq command-line tool:

bq show –format=prettyjson –job=true JOB_ID

In the query job statistics, inspect statistics.query.objectStorageStats. Relevant fields include cloudProvider, cacheBytesRead and objectStorageBytesRead. The first identifies the remote provider; the latter two distinguish bytes served from local cache from bytes read directly from remote object storage.

For a simple diagnostic, calculate cache share = cacheBytesRead / (cacheBytesRead + objectStorageBytesRead), when the denominator is nonzero. Compare this metric across representative queries and over time. It is an observed read mix, not a direct invoice reconciliation or a guarantee that all cached bytes would otherwise have incurred the same egress price.

Data platform engineer monitoring BigQuery cacheBytesRead, objectStorageBytesRead and cross-cloud query performance metrics

Track three layers together: technical behavior (cache bytes and remote bytes), user experience (p50/p95 query latency and dashboard refresh time), and economics (remote-provider egress charges plus BigQuery processing and connectivity costs). If only one layer improves, the business case may still be weak.

Five workload patterns worth testing

1. Recurring dashboards

Daily or hourly dashboards that scan stable historical partitions may repeatedly touch the same blocks. Test representative filter selections, not only a single canned query.

2. Shared curated data

Several analyst teams querying the same remote fact tables may benefit from overlap even when their SQL differs. Examine actual file-block reuse, not just table names.

3. Large one-off exploration

A first-time investigation that reads new files is unlikely to gain much from a cold cache. Repeated exploration of the same period may improve.

4. Rapidly changing source files

If files are frequently rewritten or replaced, useful cache reuse may decline. Benchmark realistic update cadence and check correctness rather than assuming a fixed cache lifetime.

5. Highly selective workloads

Good partition pruning and column selection can already minimize remote reads. Optimize query design before claiming caching savings; otherwise a cache may hide inefficient scanning.

Caching versus replication versus materialization

Federation with caching works best when the source should remain authoritative, access patterns overlap, and users can tolerate the latency and operational dependencies of remote access. It can reduce unnecessary duplicate pipelines while preserving flexible exploration.

Replicating selected tables into BigQuery may be preferable for high-volume, predictable reporting, strict latency targets, complex transformations or durable historical snapshots. Replication has its own ingestion, storage, orchestration and governance costs, but it also offers greater control over refresh cadence and data models.

Materialized reporting datasets are often the middle ground. Keep detailed data in its source cloud, query it selectively for exploration, and materialize approved KPI tables or aggregates for dashboards. This limits repeated remote scans while giving business users stable metric definitions.

For organizations reviewing this choice, Actiknow’s business intelligence consulting and data integration services describe a relevant approach to connecting multiple sources and designing reporting architecture. The engineering question is not which feature is newest; it is which architecture meets measurable decision, cost and freshness requirements.

Cloud data architects comparing cached federation, data replication and materialized reporting datasets for analytics workloads

Security, residency and governance cannot be afterthoughts

Caching stores temporary remote file blocks in the Google Cloud query region. This creates a data-residency consideration even if no permanent reporting table is created. Confirm whether the selected region is allowed for the data classification and whether contracts permit that temporary storage.

Google’s documentation says cached blocks use Google-managed encryption keys; customer-managed encryption keys are not supported for this cache. When the organization enforces the restrictNonCmekServices policy, caching may be bypassed for affected queries. Test this explicitly: an enabled configuration does not guarantee that every query uses the cache.

The service verifies metadata and access permissions against the remote source before serving cached data. Even so, teams should validate revocation behavior, identity propagation, audit trails and sensitive-column controls in their own environment. Cached reads do not remove the need for appropriate permissions on remote catalogs and objects.

Cloud security and governance specialists reviewing data residency, encryption and access controls for bigquery lakehouse caching

Operational dependencies remain. Remote catalog availability, credential rotation, metadata refresh and provider incidents can affect federation. A cache is not a disaster-recovery replica or an authoritative snapshot.

A disciplined four-week evaluation

Week 1: Baseline

Select three to five real workloads with different repetition patterns. Record remote bytes, query latency, user concurrency, provider egress charges, BigQuery compute and source-update cadence. Identify data-residency restrictions before enabling anything.

Week 2: Controlled enablement

Enable caching on an eligible connection or catalog in a nonproduction or limited-scope environment. Keep query logic, time window and workload mix comparable. Test cold and warm runs separately; include multiple query shapes against overlapping files.

Week 3: Validate economics and correctness

Compare cacheBytesRead and objectStorageBytesRead, remote provider bills and end-to-end latency. Re-run queries after source updates and permission changes. Confirm that the results match the expected source state. Investigate outliers instead of reporting only averages.

Week 4: Decide

Choose among cached federation, uncached federation, selective replication and materialized aggregates. Document an owner, rollout conditions, monitoring thresholds, fallback approach and a date to reassess as usage changes.

An executive scorecard should answer six questions:

  • What fraction of remote reads is avoided?
  • What is the change in billed transfer?
  • Did p95 latency improve?
  • Are results current and correct?
  • Is the temporary storage region compliant?
  • What happens when the remote source is unavailable?

Mistakes that undermine the business case

Treating all queries as repeated. A dashboard that changes date partitions every day may not reuse as many blocks as its schedule suggests.

Confusing cache share with financial savings. A large cacheBytesRead figure is useful evidence, but only billing and workload comparisons can establish monetary impact.

Ignoring query pruning. Reading fewer remote files through better partition design can save money regardless of cache configuration.

Comparing a warm cache to an unoptimized cold baseline. Use controlled tests and disclose warm-up conditions.

Assuming caching eliminates ETL. It avoids some remote re-reads; it does not create conformed dimensions, historical snapshots, reconciled KPIs or data-quality guarantees.

Overlooking policy constraints. Residency and encryption requirements can make a technically attractive option unsuitable.

Frequently asked questions

Is BigQuery Lakehouse caching generally available?

Yes. Google Cloud’s November 2, 2026 release notes announce general availability for cross-cloud caching. It was initially announced in Preview on August 31, 2026.

Is caching enabled automatically?

No. The documented GA feature is disabled by default and requires opt-in on a supported federated catalog or cross-cloud connection.

Does every repeated query avoid egress?

No. Benefits depend on reuse of the same file blocks, cache retention, source changes and supported configuration. A repeated query may still read remote blocks.

Does caching work only for identical SQL?

No. Different queries may benefit when they access the same remote file blocks.

Is this the same as BigQuery query-result caching?

No. Lakehouse caching operates on remote file blocks during query execution, while result caching concerns eligible completed query results.

How can teams verify cache hits?

Inspect objectStorageStats in BigQuery job statistics, particularly cacheBytesRead and objectStorageBytesRead, and compare against provider billing.

Will caching improve every query’s latency?

No. Cold reads, small workloads, low block reuse, remote metadata operations and other execution bottlenecks can limit improvement.

Does the cache preserve data residency?

Not automatically. Cached blocks reside in the Google Cloud query region; the organization must confirm that this location is permitted.

Can customer-managed encryption keys be used?

Google’s current documentation says CMEK is not supported for cross-cloud caching. Relevant organization policies can cause cache bypass.

When should we replicate instead?

Consider replication or materialized reporting datasets when predictable low latency, stable KPIs, complex transformations or durable historical snapshots outweigh the cost of maintaining copies.

Conclusion: make caching earn its place

BigQuery Lakehouse caching can reduce repeat cross-cloud reads and associated egress exposure, but the strongest architecture decision comes from workload measurements, not a feature checklist. Start with actual file-block reuse, validate bills and latency, and test security and freshness before scaling. In some environments, cached federation will be the right answer. In others, selective replication or curated reporting tables will deliver a better overall result.

If your team is weighing federation against a new ingestion pipeline, Actiknow’s BI and integration team can help scope the data sources, reporting requirements and measurable evaluation criteria. Contact Actiknow to discuss a focused architecture review.

Source notes (for editorial fact-checking)

Google Cloud BigQuery release notes, November 2, 2026: cross-cloud caching GA.

Google Cloud Lakehouse release notes, August 31, 2026: caching Preview.

Google Cloud Lakehouse documentation: About cross-cloud data access; Query remote data; objectStorageStats and cache monitoring.