qdrant/qdrant · error · CollectionError

{errors_count} of {operations_count} read operations failed{

Error message

{errors_count} of {operations_count} read operations failed{errors_separator}

What it means

Read operations fan out to the shard's replica set; when every queried replica returns an error (network, timeout, replica down), qdrant aggregates them into a service error 'N of M read operations failed' listing each replica failure. It means zero healthy replicas served that shard's read — with even one healthy replica, the read succeeds.

Source

Thrown at lib/collection/src/shards/replica_set/execute_read_operation.rs:404

            }
        }

        if responses.len() >= required_successful_results {
            Ok(responses)
        } else {
            let errors_count = errors.len();
            let operations_count = responses.len() + errors.len();
            let errors_separator = if !errors.is_empty() { ":" } else { "" };

            let mut message = format!(
                "{errors_count} of {operations_count} read operations failed{errors_separator}"
            );

            for error in errors {
                write!(&mut message, "\n  {error}").expect("writing into String always succeeds");
            }

            Err(CollectionError::service_error(message))
        }
    }
}

View on GitHub (pinned to 74f3e85b94)

Solutions

Check /cluster and telemetry: bring at least one replica per shard group back to Active state
Revive a dead/out-of-sync replica or restore the shard from a snapshot/healthy replica
Retry with backoff for transient partitions
During maintenance, never stop all replicas of the same shard at once

Defensive patterns

Strategy: retry

Try / catch

let mut attempt = 0;
loop {
    match collection.retrieve(points).await {
        Err(e) if e.to_string().contains("read operations failed") && attempt < 3 => {
            attempt += 1;
            tokio::time::sleep(Duration::from_millis(200 * attempt)).await;
        }
        other => break other,
    }
}

Prevention

Keep at least one replica per shard group running during maintenance windows
Monitor /cluster for replica states and alert before all replicas of a shard go down
Set client-side retries with backoff for read operations in distributed deployments

When it happens

Trigger: A point read (get/scroll by id) against a shard whose replicas are all stopped, unreachable, or timing out; a transient cluster-wide network partition during the read.

Common situations: Rolling restarts that stop all replicas of a shard simultaneously; misconfigured cluster addresses; replicas overloaded past request deadlines.

Related errors

AI-assisted analysis of qdrant/qdrant@74f3e85b94 (2026-08-22). Data as JSON: /api/errors/1d58d3f5bcae98cf. Report an issue: GitHub.