Skip to content

Compaction

This topic gives some information about compaction.

In NebulaGraph, Compaction is the most important background process and has an important effect on performance.

Compaction reads the data that is written on the hard disk, then re-organizes the data structure and the indexes, and then writes back to the hard disk. The read performance can increase by times after compaction. Thus, to get high read performance, trigger compaction (full compaction) manually when writing a large amount of data into Nebula Graph.

Note

Note that compaction leads to long-time hard disk IO. We suggest that users do compaction during off-peak hours (for example, early morning).

NebulaGraph has two types of compaction: automatic compaction and full compaction.

Automatic compaction

Automatic compaction is automatically triggered when the system reads data, writes data, or the system restarts. The read performance can increase in a short time. Automatic compaction is enabled by default. But once triggered during peak hours, it can cause unexpected IO occupancy that has an unwanted effect on the performance.

Full compaction

Full compaction enables large-scale background operations for a graph space such as merging files, deleting the data expired by TTL. This operation needs to be initiated manually. Use the following statements to enable full compaction:

Note

We recommend you to do the full compaction during off-peak hours because full compaction has a lot of IO operations.

nebula> USE <your_graph_space>;
nebula> SUBMIT JOB COMPACT;

The preceding statement returns the job ID. To show the compaction progress, use the following statement:

nebula> SHOW JOB <job_id>;

Operation suggestions

These are some operation suggestions to keep Nebula Graph performing well.

  • After data import is done, run SUBMIT JOB COMPACT.
  • Run SUBMIT JOB COMPACT periodically during off-peak hours (e.g. early morning).
  • To control the write traffic limitation for compactions, set the following parameter in the nebula-storaged.conf configuration file.

    Note

    This parameter limits the rate of all writes including normal writes and compaction writes.

    # Limit the write rate to 20MB/s.
    --rocksdb_rate_limit=20 (in MB/s)
    

FAQ

By default, the logs are stored under the LOG file in the /usr/local/nebula/data/storage/nebula/{1}/data/ directory, or similar to LOG.old.1625797988509303. You can find the following content.

** Compaction Stats [default] **
Level    Files   Size     Score Read(GB)  Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) Moved(GB) W-Amp Rd(MB/s) Wr(MB/s) Comp(sec) CompMergeCPU(sec) Comp(cnt) Avg(sec) KeyIn KeyDrop
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------
  L0      2/0    2.46 KB   0.5      0.0     0.0      0.0       0.0      0.0       0.0   1.0      0.0      0.0      0.53              0.51         2    0.264       0      0
 Sum      2/0    2.46 KB   0.0      0.0     0.0      0.0       0.0      0.0       0.0   1.0      0.0      0.0      0.53              0.51         2    0.264       0      0
 Int      0/0    0.00 KB   0.0      0.0     0.0      0.0       0.0      0.0       0.0   0.0      0.0      0.0      0.00              0.00         0    0.000       0      0

If the number of L0 files is large, the read performance will be greatly affected and compaction can be triggered.

"Can I do full compactions for multiple graph spaces at the same time?"

Yes, you can. But the IO is much larger at this time and the efficiency may be affected.

"How much time does it take for full compactions?"

When rocksdb_rate_limit is set to 20, you can estimate the full compaction time by dividing the hard disk usage by the rocksdb_rate_limit. If you do not set the rocksdb_rate_limit value, the empirical value is around 50 MB/s.

"Can I modify --rocksdb_rate_limit dynamically?"

No, you cannot.

"Can I stop a full compaction after it starts?"

No, you cannot. When you start a full compaction, you have to wait till it is done. This is the limitation of RocksDB.


Last update: February 1, 2023