EFS write consistency on write failure

According to the EFS docs, "Applications that perform synchronous data access and perform non-appending writes have read-after-write consistency for data access." But, what happens if the write request fails? Is it correct to assume the write operation will complete fully or not at all? Or, can the file be left in a corrupt state? My write operations will be to a specific offset, of size 4K or less. All processes (Lambdas) will obtain a write lock.