There's nothing wrong with a special-purpose tool for building approximate graphs, but calling it a "time-series database" or even quoting "inserts-per-second" is intellectually dishonest.
Many SSDs only write 4kb blocks, and writing a 64bit datapoint uncompressed to disk would not only be slow, but it would result in write amplification and wear out the disk sooner. The solution that many TSDBs, including Prometheus and Influx, involves in-memory batching with a backing WAL log file. If the in-memory batch or WAL log is lost, you would lose data as well.