- unique-postprocessing was destroying results from other
postprocessings; removed cross-updates as they had been not necessary
- unique-postprocessing did not restrict on same protocol
- inefficient concurrent update cache was redesigned completely
- increased limits for concurrent blocking queues to prevent early
time-out
This organizes all urls to be loaded in separate queues for each host.
Each host separates the crawl depth into it's own queue. The primary
rule for urls taken from any queue is, that the crawl depth is minimal.
This produces a crawl depth which is identical to the clickdepth.
Furthermorem the crawl is able to create a much better balancing over
all hosts which is fair to all hosts that are in the queue.
This process will create a very large number of files for wide crawls in
the QUEUES folder: for each host a directory, for each crawl depth a
file inside the directory. A crawl with maxdepth = 4 will be able to
create 10.000s of files. To be able to use that many file readers, it
was necessary to implement a new index data structure which opens the
file only if an access is wanted (OnDemandOpenFileIndex). The usage of
such on-demand file reader shall prevent that the number of file
pointers is over the system limit, which is usually about 10.000 open
files. Some parts of YaCy had to be adopted to handle the crawl depth
number correctly. The logging and the IndexCreateQueues servlet had to
be adopted to show the crawl queues differently, because the host name
is attached to the port on the host to differentiate between http,
https, and ftp services.
- added order option to solr queries to be able to retrieve document
lists in specific order, here: link length
- added HyperlinkEdge class which manages the link structure
- integrated the HyperlinkEdge class into clickdepth computation
- extended the linkstructure.json servlet to show also the clickdepth
and other statistic information
instead of TreeMaps)
- enhanced memory footprint of database indexes (by introduction of
optimize calls)
- optimize calls shrink the amount of used memory for index sets if they
are not changed afterwards any more
different from normal requests. This happens if the remote solr is
actually a solrCloud; in such cases the luke request returns only the
result of the single solr peer, not the whole cloud.
also done: some refactoring.
url along with the load date. While this takes much more memory, it
eliminates database lookups for getURL() requests, which happen equally
often. This speeds up remote solr configurations.
The resource observer is now able to recognize free disk space AND
available space for YaCy. The amount of space which is assigned for YaCy
are defined in new settings in the configuration file.
Furthermore, there is now a cleanup process which deletes files in case
that an autodelete is activated. The autodelete is now BY DEFAULT ON if
the disk space is low, which means that YaCy starts to delete documents
when the disk is full!
- redesigned the instance mirror class (which was a mess)
- added final method to close a searcher (which otherwise keeps a cache)
- changed cache clear method which iterates over resources and calls
clear to all caches in the searcher resources
- refactored all code which uses URIMetadataRow as standard for word
hash length and word hash ordering and moved that to the class 'Word',
becuase the class URIMetadataRow defined the old metadata data structure
and should be superfluous in the future
- removed unused methods from URIMetadataRow as preparation for further
removal of that class
- all non-dht targets (previously separated into 'robinson' for dht-like
queries and 'node' for solr queries) are non 'extra' peers, which are
queries using solr
- these extra-peers are now selected using a ranking on last-seen,
peer-tag-matches, node-peer flags, peer age, and link count. The ranking
is done using a weight and a random factor.
- the number of extra peers is 50% of the dht peers
- the dht peers now exclude too young peers to prevent bad results
during strong growth of the network
- the number of dht peers (and therefore extra-peers) is reduced when
the memory of the peer is low and/or some documents still appear in the
indexing-queue. This shall prevent a peer from deadlocks when p2p
queries are made in a fast sequence on weak hardware.
causes Solr error (and wordindex likely finds suggestion)
org.apache.solr.core.SolrCore org.apache.solr.common.SolrException: org.apache.solr.search.SyntaxError: Cannot parse 'text_t:""d"': Lexical error at line 1, column 12. Encountered: <EOF> after : ""
at org.apache.solr.handler.component.QueryComponent.prepare(QueryComponent.java:171)
at org.apache.solr.handler.component.SearchHandler.handleRequestBody(SearchHandler.java:187)
at org.apache.solr.handler.RequestHandlerBase.handleRequest(RequestHandlerBase.java:135)
at net.yacy.cora.federate.solr.connector.EmbeddedSolrConnector.query(EmbeddedSolrConnector.java:179)
at net.yacy.cora.federate.solr.connector.EmbeddedSolrConnector$DocListSearcher.<init>(EmbeddedSolrConnector.java:345)
at net.yacy.cora.federate.solr.connector.EmbeddedSolrConnector.getCountByQuery(EmbeddedSolrConnector.java:364)
at net.yacy.cora.federate.solr.connector.MirrorSolrConnector.getCountByQuery(MirrorSolrConnector.java:326)
at net.yacy.cora.federate.solr.connector.ConcurrentUpdateSolrConnector.getCountByQuery(ConcurrentUpdateSolrConnector.java:440)
at net.yacy.search.index.Segment.getWordCountGuess(Segment.java:464)
at net.yacy.data.DidYouMean.getSuggestions(DidYouMean.java:181)
at suggest.respond(suggest.java:73)
- the admin user name can be configured, in apiExec calls the default "admin" username is used.
TODO: the bin/apicall.sh script should likely take that into account.
as path for solr index dumps (instead of the SEGMENTS path). This will
make a maintenance of index backups easier. It will also provide a tool
to migrate from an freeworld index to a webportal index.
work around the unfolding process in Solr's BinaryResponseWriter.
This was a huge performance bottleneck in the embedded solr connector
and the problem is actually on Solr side, but we have now a workaround.
- This made it possible to abstract a high-performance index access
method which is implemented as method getDocumentListByParams. That
method is also implemented in the SolrServerConnector and provides a
very efficient access to a solr index if the index is embedded.
- a popular use of the document list retrieval is a result count which
can now also make use of the new method, via getDocumentCountByParams.
- enhanced the Error cache which now does not store error documents
within the ram cache if the document is also written to solr. When
documents are retrieved from the cache, they are partly read from the
ram cache and if not existent there, from the Solr index.
servlet since YaCy 1.63. This is much more performant for the client
than using the XMLResponseWriter because parsing of XML data is very CPU
intensive. Older YaCy peers are still requested using the
XMLResponseWriter but the majority of YaCy peers already respond with
the binary writer. This makes remote searches much faster and less CPU
intensive.
not-flushed Solr cache is now handled in this way:
- it is smaller by default
- an Solr-internal process is started to flush the cache periodically
(this does NOT clean the cache, just removes old objects)
- a Solr-external process (the standard YaCy cleanup-process) now has
direct access to the solr internal cache and flushes them completely.
The time frame for such a flush is defined by the cleanup-process
frequency, by default 10 minutes.
the embedded Solr (the default). This was obtained by cirumventing solrj
search encapsulation and the implementation of direct index access
methods to Solr.
The effect will not only be seen during search, but this has also a
strong effect on suggestions (much more) and less CPU power usage during
index distribution (which needs many search requests)
which had a problem because of badly used concurrency.
This fix also caused a redesign of the whole host deletion process.
This should fix bug http://bugs.yacy.net/view.php?id=250
This shall fulfill the following requirement:
If a document A links to B and B contains a 'canonical C', then the
citation rank computation shall consider that A links to C and B does
not link to C.
To do so, we first must collect all canonical links, find all references
to them, get the anchor list of the documents and patch the citation
reference of these links.
webgraph index which is temporary filled with the crawl profile key.
This is used to select a set of documents for post-processing as soon as
a crawl is finished. Now the postprocessing for a specific crawl is
started when that specific crawl is finished and not at the end of all
post-processing steps.
there for deletion), this fixes a problem for the deletion of old
documents for new crawl starts
- added clickdepth and citation computation for fail documents
- replaced load failure logging by information which is stored in Solr
- fixed a bug with crawling of feeds: added must-match pattern
application to feed urls to filter out such urls which shall not be in a
wanted domain
- delegatedURLs, which also used ZURLs are now temporary objects in
memory
for anchor attributes.
- this caused that large portions of the parser code had to be adopted
as well
- added a counter target_order_i for anchor links in webgraph
computation
all unique links! This made it necessary, that a large portion of the
parser and link processing classes must be adopted to carry a different
type of link collection which carry a property attribute which are
attached to web anchors.
- introduction of a new URL class, AnchorURL
- the other url classes, DigestURI and MultiProtocolURI had been renamed
and refactored to fit into a new document package schema, document.id
- cleanup of net.yacy.cora.document package and refactoring
fuzzy_signature_copycount_i, which count the number of copies of
non-unique documents and assigns this to each document. Thus, each
document there is a number assigned which shows how many copies of this
document exists.
These fields are disabled by default.
regular expression on th url: the collection attribut for a crawl start
may be now either a token or a list of tokens, seperated by ',' where a
token is either a string or a pair <string,pattern> where the string is
separated to the pattern with a ':' and the string is assigned to the
document as collection only if the pattern matches with the url.
in intranets and the internet can now choose to appear as Googlebot.
This is an essential necessity to be able to compete in the field of
commercial search appliances, since most web pages are these days
optimized only for Google and no other search platform any more. All
commercial search engine providers have a built-in fake-Google User
Agent to be able to get the same search index as Google can do. Without
the resistance against obeying to robots.txt in this case, no
competition is possible any more. YaCy will always obey the robots.txt
when it is used for crawling the web in a peer-to-peer network, but to
establish a Search Appliance (like a Google Search Appliance, GSA) it is
necessary to be able to behave exactly like a Google crawler.
With this change, you will be able to switch the user agent when portal
or intranet mode is selected on per-crawl-start basis. Every crawl start
can have a different user agent.
jdk-based logger tend to block
at java.util.logging.Logger.log(Logger.java:476) in concurrent
environments. This makes logging a main performance issue. To overcome
this problem, this is a add-on to jdk logging to put log entries on a
concurrent message queue and log the messages one by one using a
separate process.
- FTPClient uses the concurrent logging instead of the log4j logger
references_internal_url_sxt because they had been shown to be
superfluous. The citation of referrer in the host browser is possible
without them. Therefore now the host browser does not only show
internal, but also external referrer to each link.
While the values for the reference evaluation are computed, also a
backlink-structure can be discovered and written to the index as well.
The host browser has been extended to show such backlinks to each
presented links. The host browser therefore can now show an information
where an document is linked. The new citation reference is computed as
likelyhood for a random click path with recursive usage of previously
computed likelyhood. This process is repeated until the likelyhood
converges to a specific number. This number is then normalized to a
ranking value CRn, 0<=CRn<=1. The value CRn can therefore be used to
rank popularity within intra-domain link structures.
- move setting of system property solr.directoryFactory=solr.MMapDirectoryFactory to solrcore.properties
- add check of os.arch for 64bit system, if it fails use default/solrcore.x86.properties (if exists) as solrcore.properties
reason: on 32bit MMapDirectoryFactory may fail with.....
Caused by: java.io.IOException: Map failed
at sun.nio.ch.FileChannelImpl.map(FileChannelImpl.java:849)
at org.apache.lucene.store.MMapDirectory.map(MMapDirectory.java:283)
soft commits, reduced caching size of search events, ensured that solr
results are processed before connection is closed to keep that stuff not
too long in RAM
API recording for this method so it can be repeated automatically. The
index dump generation is now also available for API recording. Added
some synchronization in backend which was necessary for this.
id to be tested, but with a collection of ids. This will cause only a
single call to solr instead of many. The result is a much better
performace when testing the existence of many urls. The effect should
cause very much less IO during index transmission, both on sender and
receiver side.
- removed httpclient 3.1 which has been used by solrj < 4.x.x and is now
not used any more
- fixed some parts in YaCy which used methods from httpclient 3.1
Because the index size is now provided by solr, and the only way to do
that is a match for [* TO *], a size computation is quite complex and
time-consuming. Therefore this patch prevents that the method is called
at all and if necessary puts a DOS-preventing barrier in front of it.
- check geolocation coordinates and accept only those, which are
well-formed
- the solr push process does not stop crawling any more if after 20
requests to Solr Solr does not accept the record. Instead, a severe log
entry asks the user to create a bug request
- intruduced raw-queries for the re-introduced byId-Queries (they are
hopefully faster than full edismax queries)
- removed the cached solr connector (testing this) to rely only on the
solr built-in search caches. That should save some RAM (also). We will
see if this is usable.
adjusted to smaller and 1-core devices.
- the workflow processor now starts no process at all. these are started
as soon as parser/condenser/indexing queues are filled.
- better abstraction
holds the number of documents for the host where the document is hosted.
This is necessary for ranking and the norming of references per local
host in the ranking computation.
references_external_i and references_exthosts_i. These can be used to
count and evaluate the number of external links to every web page. An
experimental ranking function can be i.e.:
div(add(references_internal_i,product(references_external_i,references_exthosts_i)),add(clickdepth_i,1))
Michael Peter Christen