The disclosed technology is generally directed to IoT device update failure recovery. In one example of the technology, after writing an updated release to memory, a determination is made whether the updated release is valid. The updated release includes a plurality of image binaries. If the updated release is determined to be valid, the updated release is made the current release. A determination is made as to whether the current release is stable. Upon determining that the current release is unstable, an auto-rollback is performed. Performing the auto-rollback includes, via at least one processor, automatically: obtaining an uncompressed backup of a previous release; making the uncompressed backup of the previous release the current release; and executing the uncompressed backup.