Use block size from HDFS configuration for Large Files calculation

[pjeli]

In NNA today, particularly if you look around here:

NNAnalytics/src/main/java/org/apache/hadoop/hdfs/server/namenode/cache/SuggestionsEngine.java

Lines 161 to 165 in b17e8e6

	Collection<INode> mediumFiles =
	nnLoader.combinedFilter(
	files,
	new String[] {"fileSize", "fileSize"},
	new String[] {"lte:134217728", "gt:1048576"});

You will see that NNA uses a hardcoded cut off of 128 Megabyte block sizes to distinguish between "Medium Files" and "Large Files".

We should instead utilize the bytes count from dfs.blocksize value found in hdfs-site.xml (Configuration object in NNA, programmatically) passed into NNA that came from the source cluster.

[americanyorkie]

What about making the file size definitions user configurable here as it's reasonable to expect differing opinions on what constitutes a particular file size from users. Currently file sizes are defined:
tiny > 0 && tiny <= 1024
small > 1024 && small <= 1048576
medium > 1048576 && medium <= 134217728
large > 134217728
There would be value in being able to examine the farther end of the scale a bit more granularly. Importing tables from RDBMS' can result in files 10's & 100's of GBs in size, for example.

[akshatgit]

Good idea, maybe a web UI where the user can select different filter to sort/group the files would be a better interface

[pjeli]

Hmm, yes a good idea @americanyorkie -- something to keep in mind though is that those are cached results you see -- so while yes that is possible to change them however the result may not be reflected until the next SuggestionEngine run.

Still though; this is probably fine.

I can see an admin-only REST endpoint that would set these. For example, a naive one like: /setAttributes?tiny=1024&small=1048576&medium=134217728 (assuming then that large is anything greater than 134217728).

Thoughts?

I don't think it will be possible to have different settings per user though...
We could certainly add a "gigantic file" category too. 😆

[pjeli]

I still think this is best to be fetched from the HDFS configuration file (hdfs-site.xml) as that should be the same value used by the active NameNode. If a different value is desired then it can be changed for just the NNA hosts hdfs-site.xml.

Changing this value on the fly will not be good for NNA so it needs to be a hard value decided on bootstrap time.

[pjeli]

An additional justification is that once NNA bootstraps from a cluster NameNode (Observer or Standby), it will anyway have expected configuration.

[pjeli]

More thoughts on this one -- I think we should give a statistic by which we measure tiny, small, and medium files.

I think ratios are probably the best measure here. If we were to retain the same hardcoded values then... assuming hdfs-site.xml has a blocksize of 128MB...

Large files = Greater than blocksize
Medium files = Greater than or equal to 1/128 of blocksize but less than large files
Small files = Greater than or equal to 1/131072 of blocksize but less than medium files
Tiny files = Greater than 0 bytes but less than small files

The ratios aren't very intuitive however. Might be better to stick with the hardcoded 1KB and 1MB sizes. Just dumping thoughts.