This is a question related to this question: Hive/Hadoop intermittent failure: Unable to move source to destination
We found that we could avoid the problem of “Unable to move source … Filesystem closed” by setting fs.hdfs.impl.disable.cache
to true
However, we also observed that the SparkSQL queries became very slow — queries that used to finish within a few seconds now take more than 30 to 40 seconds to finish (even when the query is very simple, like reading a tiny table).
Is this normal?
My understanding of fs.hdfs.impl.disable.cache
being true means that FileSystem#get()
would always createFileSystem()
instead of returning a cached FileSystem
. This setting prevents a FileSystem
object from being shared by multiple clients and it really makes sense, because it would prevent, for example, two callers of FileSystem#get()
from closing each other’s filesystem.
(For example, see this discussion )
This setting would slow things down, but probably not by so much.
From: hadoop-source-reading
/** * Returns the FileSystem for this URI's scheme and authority. The scheme of * the URI determines a configuration property name, * <tt>fs.<i>scheme</i>.class</tt> whose value names the FileSystem class. * The entire URI is passed to the FileSystem instance's initialize method. */ public static FileSystem get(URI uri, Configuration conf) throws IOException { String scheme = uri.getScheme(); String authority = uri.getAuthority(); if (scheme == null) { // no scheme: use default FS return get(conf); } if (authority == null) { // no authority URI defaultUri = getDefaultUri(conf); if (scheme.equals(defaultUri.getScheme()) // if scheme matches // default && defaultUri.getAuthority() != null) { // & default has // authority return get(defaultUri, conf); // return default } } String disableCacheName = String.format("fs.%s.impl.disable.cache", scheme); if (conf.getBoolean(disableCacheName, false)) { return createFileSystem(uri, conf); } return CACHE.get(uri, conf); }
Would the slowness point to some other networking issues, such as resolving domain names? Any insights to this problem are welcome.