pebble: Performance degradation on massive writing

Hello, we are importing large amount of data (261,000,000 records) into the pebble database - we write them using Put operation. The performance of writes degrades and new writes takes more and more time. From the metrics table we see the level 0 becomes huge and it is probably the cause.

• Closing and reopening the database does not help
• Enclosing writes into batches did not help
• When we run db.Compact([]byte{0x00}, []byte{0xFF}) in the middle of the import, it becomes fast again
• Without compact, the import takes 25 hours
• When we run Compact ± every 1 minute, it is finished in 3 hours (the Compacts time included)

I believe this is a performance bug of pebble. Is there something we should check?

Thanks

Metrics (without calling Compact manually) - the level 0 grows to hundreds gigabytes:

__level_____count____size___score______in__ingest(sz_cnt)____move(sz_cnt)___write(sz_cnt)____read___r-amp___w-amp
    WAL         2   6.3 M       -   401 M       -       -       -       -   401 M       -       -       -     1.0
      0    387143   267 G    1.08   395 M     0 B       0     0 B       0   263 M    17 K     0 B       4     0.7
      1        25    89 M    1.35   570 M     0 B       0     0 B       0   2.3 G     804   2.3 G       1     4.2
      2       135   598 M    1.04   393 M     0 B       0   144 M      78   1.6 G     258   1.6 G       1     4.2
      3       414   5.1 G    1.00   494 M     0 B       0    10 M       4   1.8 G     155   1.8 G       1     3.8
      4      1841    45 G    0.97   368 M     0 B       0   4.0 M       1   1.4 G      60   1.4 G       1     4.0
      5     16340   397 G    0.95     0 B     0 B       0     0 B       0     0 B       0     0 B       1     0.0
      6     31725   3.4 T       -     0 B     0 B       0     0 B       0     0 B       0     0 B       1     0.0
  total    437623   4.1 T       -   401 M     0 B       0   158 M      83   7.8 G    18 K   7.2 G      10    20.0
  flush       112
compact       431   3.2 T   5.1 M       1          (size == estimated-debt, score = in-progress-bytes, in = num-in-progress)
  ctype       348       0       0      83       0       0  (default, delete, elision, move, read, rewrite)
 memtbl         2   8.0 M
zmemtbl         0     0 B
   ztbl         0     0 B
 bcache         0     0 B   59.8%  (score == hit-rate)
 tcache     1.9 K   1.2 M   89.8%  (score == hit-rate)
  snaps         1       - 17835168338  (score == earliest seq num)
 titers         0
 filter         -       -    0.0%  (score == utility)