Preserve remote-only initiator reads

When object_storage_remote_initiator=1 and only object_storage_remote_initiator_cluster is set on the initiator, this condition now forces fallback_to_pure=false even though there is no local object_storage_cluster to defer. read then takes the clustered path and rewrites the remote query with make_cluster_function=true, producing s3Cluster('', ...) instead of the previous s3(...); on remote users whose profile defines object_storage_cluster, that bypasses the fallback table function and the read runs only on the remote initiator, regressing the existing test_object_storage_remote_initiator_without_cluster_function case that expects replica subqueries. Only defer local cluster resolution when cluster_name_from_settings is non-empty.

Useful? React with 👍 / 👎.

Restrict fallback to object_storage_cluster setting

Because this check is in the IStorageCluster base and only looks at the resolved cluster name, enabling object_storage_cluster_fallback_if_empty also changes explicit cluster table functions such as s3Cluster('typo', ...): getClusterName returns the function argument, tryGetCluster returns null, and the read falls back to pure_storage on the initiator instead of reporting the bad cluster. This makes a misspelled or temporarily empty explicit cluster silently run as a single-node read whenever the setting is enabled, so the fallback needs to be gated to names that actually came from the object_storage_cluster setting or otherwise kept out of explicit *Cluster calls.

Useful? React with 👍 / 👎.

Default remote initiator to object_storage_cluster

When object_storage_remote_initiator=1 is used with alternative syntax and only object_storage_cluster is set, this new branch leaves remote_initiator_cluster unset because object_storage_remote_initiator_cluster is empty. That was a supported path: the old code defaulted the remote initiator cluster to cluster_name_from_settings, but now getQueryProcessingStage/read either throw BAD_ARGUMENTS or, with object_storage_cluster_fallback_if_empty=1, run the read locally instead of on a remote initiator. Please preserve the previous default to the object-storage cluster when the explicit remote-initiator cluster setting is absent.

Useful? React with 👍 / 👎.

Apply fallback to table-engine object_storage_cluster

This disables object_storage_cluster_fallback_if_empty for persistent S3/Iceberg tables created with SETTINGS object_storage_cluster='...', because those tables store that setting in getOriginalClusterName while cluster_name_in_settings remains false. As a result, SELECT ... SETTINGS object_storage_cluster_fallback_if_empty=1 against such a table still resolves the missing/empty cluster normally and throws instead of falling back to the local storage, even though the new setting is described as applying when object_storage_cluster is set. Please distinguish table-engine settings from explicit *Cluster arguments before rejecting the fallback.

Useful? React with 👍 / 👎.

Maybe rename it to object_storage_cluster_fallback_to_local_if_empty

It seems like I don't understand this PR enough. From the description, I got the idea that if the remote cluster isn't available (i.e, no nodes alive), the setting would allow to execute it on this machine.

But the method useObjectStorageClusterFallbackIfEmpty, which indicates it is related to that setting, doesn't even check that setting.

Agree, I'll change it on more clean code.

Do not ignore a missing remote initiator cluster

When object_storage_remote_initiator_cluster is supplied but misspelled or its swarm is empty, this allow_null path makes getClusterImpl return nullptr; the read path then falls through to readFallBackToPure whenever object_storage_cluster_fallback_to_local_if_empty is enabled. That silently runs a query that explicitly requested a remote initiator on the local server instead of reporting the bad initiator cluster, while the new fallback is documented for an empty or unknown object_storage_cluster, not for a missing initiator cluster.

Useful? React with 👍 / 👎.

Preserve *Cluster rewrite unless fallback is needed

With object_storage_remote_initiator=1, object_storage_remote_initiator_cluster set, and a valid non-empty object_storage_cluster, this condition forces the read into the pure remote path even when object_storage_cluster_fallback_to_local_if_empty is false. That bypasses the later make_cluster_function=true rewrite to s3Cluster/icebergCluster that this file still documents as needed for old ClickHouse versions, so during rolling upgrades the remote initiator can receive s3(... SETTINGS object_storage_cluster=...) instead of the compatible *Cluster call and fail to parse or distribute it. Please keep the *Cluster path unless the cluster is actually missing/empty and fallback is requested.

Useful? React with 👍 / 👎.

Thanks for the review. The observation is correct that with object_storage_remote_initiator=1 and a non-empty object_storage_remote_initiator_cluster, setting-syntax reads (s3() / ENGINE ... SETTINGS object_storage_cluster=...) always take the pure remote path and send plain s3()/iceberg() rather than rewriting to *Cluster.

That is intentional and should stay.

Why not restore *Cluster when fallback is off

Under remote-initiator deferral, the initiator must not assume it knows how to resolve object_storage_cluster:

• OSC may be unknown / unset on the local node and defined only on the remote initiator (or in the remote user’s profile).
• Or the reverse: OSC is set on the initiator query, but the remote swarm is empty/unknown and should apply object_storage_cluster_fallback_to_local_if_empty.

Sending s3Cluster('name', ...) (or equivalent) bakes the cluster name into the table-function argument. Explicit *Cluster never applies object_storage_cluster_fallback_to_local_if_empty, and it also forces the initiator to pick a concrete cluster-argument form even when the name is only meaningful on the remote side.

What we send instead

SELECT ... FROM remote('host', s3(...))
SETTINGS object_storage_cluster = '...', object_storage_cluster_fallback_to_local_if_empty = ...

• Keep the plain s3() / iceberg() function.
• Pass object_storage_cluster as a query setting (not a function argument).
• If local OSC is empty, omit it so the remote can supply its own.

That lets the remote initiator distribute when it knows the cluster, or fall back locally when the setting allows — without the local node needing a live/known swarm definition.

The make_cluster_function=true / “old ClickHouse versions” rewrite remains for the non-deferred path (e.g. object_storage_remote_initiator=1 with only object_storage_cluster, no object_storage_remote_initiator_cluster), where the initiator cluster is the object-storage cluster itself and a *Cluster rewrite is appropriate.

So we will not change the deferral path back to *Cluster based on this comment.