Before moving the archive, ask who keeps reading it

The archive is suddenly busy every morning. Perhaps a new training job needs its contents. Perhaps a pipeline keeps fetching a copy it already has. Google's Storage Intelligence update helps reveal that pattern and preview a bulk change, but those two explanations call for different responses. One may justify moving the data; the other may justify fixing the job. Useful work might need no interruption at all.

The September 25, 2026 announcement introduced general availability for the advisor and expanded batch operations for Storage Intelligence customers. Daily snapshots surface unusual cold-storage operations, rate-limit errors and data leaving for another region. They describe usage patterns over time, rather than providing an immediate verdict on whether a service is healthy.

Follow the morning read to its job

Imagine a hypothetical evaluation job that has started reading archived training data every day. Objects used that often may suit a different storage class, with different access costs. But if the daily download serves no purpose, removing it could be better than moving everything it touches. The storage finding is the same; the job explains what to do about it.

Bring its maintainer the bucket, the period and the matching job activity. “Does this job need these objects every day?” gives them something concrete to answer. “Why did the bill rise?” asks them to begin the investigation again. They may identify a mistake, or explain why those reads are doing exactly what the team wants.

A daily snapshot helps reveal repetition. It cannot tell you that the job is healthy right now or that an object has stayed unchanged since the snapshot. Follow the finding into the job and storage records before choosing the current objects a change should touch.

Look inside the dry-run count

A dry run previews a batch without applying its changes. The batch documentation describes object counts and possible errors, plus total size when selecting by prefix, the beginning of an object's name. Listing objects can still cost money. The preview answers what matches your rule; it does not explain why every match belongs in the change.

A CSV manifest, a list of objects to process, can specify a generation identifying an object's stored version. Leave that out and the job processes the live version. Leave a prefix list empty and it selects the whole bucket. A reasonable-looking count can therefore hide objects newer, or far more numerous, than the reviewer intended.

Save the requested operation alongside its exact selection. In the bounded-action approach, approval covers that reviewed manifest and configuration. Alter either and there is a different batch to review. Keep the submitted job comparable to the one the owner actually approved.

The manifest gives the reviewer individual items to inspect; a generation says which stored version is meant. That distinction becomes useful while a bucket changes between investigation and execution. Recognizing the object name does not establish that its contents are still the same.

Suppose archive/ contains both obsolete output and the evaluation datasets the new job still needs. A prefix count includes both groups, even if only one should move to another storage class. Ask the maintainer to check the matched objects before submitting that move. The folder-like name is a convenient selector, not a description of each object's purpose.

Find out which objects actually changed

After the job runs, inspect its successes and failures before launching another broad batch. Automatic retries are supported, and partial progress can mean real changes have already happened. Use the results to identify the unfinished work rather than making successful items travel through the same operation again.

Deletion is a separate decision with a separate recovery check against the bucket's actual settings. Keep required data when restoration is unproven. For the dataset move, come back to those morning reads: the maintainer should recognize the needed objects, the submitted selection should match the reviewed one, and the results should say what moved and what remains. That is how a storage finding becomes an understood change, rather than just a smaller bill with a new mystery attached.