It seems from the comments that the data is gone because it resided in an instance that was part of a managed instance group. When the group required additional resources, a new instance was created based on the instance template. Presumably, the template did not contain any data that had been gathered on a given instance, thus the new instance contained no data. This is precisely how it is possible to lose the data stored on these instances.
In addition, storing data directly on these instances will almost immediately lead to data discrepancies if there's no system in place to synchronize data across instances. Given a group with 2 instances (A and B), a load balancer sends a request to instance A resulting in a write. Moments later, the load balancer sends a new request to instance B seeking out the data that was recently written. It will find nothing. Thus, the data must either be synchronized which can be very costly and error prone, or better yet migrated outside the scaling instance group.
Stateless instance design
In a GCE Discussion group post about scaling single VMs, Kamran describes how to prevent this sort of data loss/discrepancy. With instances having no persistent state (continuing to exist after they've been shut down), they can be considered stateless. The advantage to making all instances in a given group stateless is that one can expect any instance in the group to behave the same way and handle tasks the same way. This, therefore lends itself quite well to horizontal scaling (having more instances to process more work rather than simply having a more powerful instance).
Storage outside the instance group
Generally, this stateless design requires that you migrate your data/file storage outside the instance group. In the case of SQL, one could move the data to an instance of Cloud SQL accessible to the instance group or Cloud Datastore for a NoSQL alternative. For file storage, one could migrate files to Cloud Storage. You could even use Cloud Storage FUSE to mount a given GCS bucket as a file system on your managed instances.