compact/write: Don't write down empty blocks (?)

[bwplotka]

I had no samples for a while in my Prometheus (because of wrong scraping configuration?). TSDB dumped empty block into my disk which looks like this:

01C6XA7C2K2EJKV6GGWQCGNYTC/
-- index (39.53 KB)
-- meta.json (408 B)
-- chunks/
     < empty dir >

Meta file:

{
    "version": 1,
    "ulid": "01C6XA7C2K2EJKV6GGWQCGNYTC",
    "minTime": 1519236000000,
    "maxTime": 1519243200000,
    "stats": {},
    "compaction": {
        "level": 1,
        "sources": [
            "01C6XA7C2K2EJKV6GGWQCGNYTC"
        ]
    },
}

This causes some tiny problems in Thanos that we need to workaround, but my main questions are:

1. Have you seen that before? How often can that happen?
2. Is that necessary to produce those?

In my personal opinion, it can cause lot's of unnecessary compactions (and downsampling for Thanos) and index lookups. It gets even worse for Thanos or other long storage solutions, because these lookups & compactions are bit more expensive.

Do you think we can just don't write empty blocks down?

The only negative consequence I can think of is the inability to tell from the storage perspective (looking on the blocks) if the Prom/TSDB was down or just Prometheus did not scrape anything (no samples in TSDB).

[gouthamve]

Even I think that writing empty blocks can be removed.

See: #245