mirror of
https://github.com/Alfresco/alfresco-community-repo.git
synced 2025-06-30 18:15:39 +00:00
2e62d4fb29
30342: Dev branch for Site performance issues (including rework of AuthorityService.getAuthorities() to use a 'lazy' set and DM indexing rework)
ALF-9899 Huge share site migration, add group to site and user access site related performance issue.
ALF-9208 Performance issue, during load tests /share/page/user/user-sites is showing to be the most expensive.
ALF-9692 Performance: General performance of Alfresco degrades when there are 1000s of sites present
- ancestor-preloading
- hasAuthority
- huge site test
30370: - Save changed to do with adding childAuthorityCache to AuthorityDAOImpl
- Increase aspectsTransactionalCache size as it blows up
30387: Experimental solution to 'cascading reindex' performance problem
- Now only Lucene container documents for a single subtree are reprocessed on addition / removal of a secondary child association
- No need to delete and re-evaluate ALL the paths to all the nodes in the subtree - just the paths within the subtree
- Lucene deltas now store the IDs of ANCESTORs to mask out as well as documents to reindex
- Merge handles deletion of these efficiently
- Node service cycle checks changed from getPaths to recursive cycleCheck method
- Adding a group to 60,000 sites might not require all paths to all sites to be re-evaluated on every change!
30389: Missed files from last checkin
30390: Optimizations / fixes to Alan's test!
30393: Bug fix - wasn't adding new documents into the index!
30397: Fixed a problem with bulk loading trying to bulk load zero parent associations
Also tweaked reindex calls
30399: Correction - don't cascade below containers during path cascading
30400: Another optimization - no need to trigger node bulk loading during path cascading - pass false for the preload flag
30404: Further optimizations
- On creation of a secondary child association, make a decision on whether it is cheaper to cascade reindex the parent or the child, based on the number of parent associations to the child
- Assumes that if there are more than 5 parent associations, it's cheaper to cascade reindex the parent
- Add a new authority to a zone (containing 60,000 authorities) - cascade reindex the authority, not the zone
- Add a group (in 60,000 sites) to a site - cascade reindex the site, not the group
- Caching of child associations already traversed during cascade reindexing
- Site creation time much reduced!
30407: Logic fix: Use 'delete only nodes' behaviour on DM index filtering and merging, now we are managing container deletions separately
30408: Small correction related to last change.
30409: Correction to deletion reindex behaviour (no need to regenerate masked out containers)
- Site CRUD operations now all sub-second with 60,000 sites!
30410: Stop the heartbeat from trying to load and count all site groups
- Too expensive, as we might have 60,000 sites, each with 4 groups
- Now just counts the groups in the default zone (the UI visible ones)
30411: Increased lucene parameters to allow for 'path explosion'
- 9 million lucene documents in my index after creating 60,000 Share sites (most of them probably paths) resulting in sluggish index write performance
- Set lucene.indexer.mergerTargetIndexCount=8 (142 documents in smallest index)
- Increased lucene.indexer.maxDocsForInMemoryMerge, lucene.indexer.maxDocsForInMemoryIndex
30412: Test fixes
30413: Revert 'parent association batch loading' changes (as it was a bad idea and is no longer necessary!)
- Retain a few caching bug fixes however
30416: Moved UserAuthoritySet (lazy load authority set) from PermissionServiceImpl to AuthorityServiceImpl
30418: - Remove 'new' hasAuthority from authorityService so it is back to where we started.
- SiteServiceHugeTest minor changes
30421: Prevent creation of a duplicate root node on updating the root
- Use the ANCESTOR field rather than ISCONTAINER to detect a node document, as the root node is both a container and a node!
30447: Pulled new indexing behaviour into ADMLuceneIndexerImpl and restored old behaviour to AVMLuceneIndexerImpl to restore normal AVM behaviour
30448: - Cache in PermissionServiceImpl cleared if an authority container has an association added or removed
Supports the generateKey method which includes the username
Supports changes in group structures
- Moved logic to do with ROLE_GUEST from PermissionServiceImpl to AuthorityServiceImpl
30465: - Tidy up tests in SiteServiceTestHuge
30532: - Added getContainingAuthoritiesInZone to AuthorityService
- Dave Changed PeopleService.getContainerGroups to only return groups in the DEFAULT zone
- Fixed RM code to use getAuthoritiesForUser method with just the username again.
30558: Build fixes
- Fixed cycleCheck to throw a CyclicChildRelationshipException
- More tidy up of AVM / ADM indexer split
- Properly control when path generation is cascaded (not required on a full reindex or a tracker transaction)
- Support indexing of a 'fake root' parent. Ouch my head hurts!
30588: Build fixes
- StringIndexOutOfBoundsException in NodeMonitor
- Corrections to 'node only' delete behaviour
- Use the PATH field to detect non-leaf nodes (it's the only stored field with which we can recognize the root)
- Moved DOD5015Test.testVitalRecords() to the end - the only way I could work out how to get the full TestCase to run
30600: More build fixes
- Broadcast ALL node deletions to indexer (even those from cascade deletion of primary associations)
- Allows indexer to wipe out all affected documents from the delta even if some have already been flushed under different parents by an intricate DOD unit test!
- Pause FTS in DOD5015Test to prevent intermittent test failures (FTS can temporarily leave deleted documents in the index until it catches up)
- More tidy up of ADMLuceneIndexerImpl
- flushPending optimized and some unnecessary member variables removed
- correction to cascade deletion behaviour (leave behind containers of unaffected secondary references)
- unused MOVE action removed
- further legacy logic moved into AVMLuceneIndexerImpl
30620: More build fixes
- Cope with a node morphing from a 'leaf' to a container during its lifetime
- Container documents now created lazily in index as and when necessary
- Blank out 'nth sibling' field of synthesized paths
- ADMLuceneTest now passes!
- TaggingServiceImplTest also passes - more special treatment for categories
30627: Multi tenancy fixes
30629: Possible build fix - retrying transaction in ReplicationServiceIntegrationTest.tearDown()
30632: Build fix - lazy container generation after a move
30636: Build fix: authority comparisons are case sensitive, even when that authority corresponds to a user (PermissionServiceTest.testPermissionCase())
30638: Run SiteServiceTestHuge form a cmd line
set SITE_CPATH=%TOMCAT_HOME%/lib/*;%TOMCAT_HOME%/endorsed/*;%TOMCAT_HOME%/webapps/alfresco/WEB-INF/lib/*;\
%TOMCAT_HOME%/webapps/alfresco/WEB-INF/classes;%TOMCAT_HOME%/shared/classes;
java -Xmx2048m -XX:MaxPermSize=512M -classpath %SITE_CPATH% org.alfresco.repo.site.SiteServiceTestHuge ...
Usage: -Daction=usersOnly
-Dfrom=<fromSiteId> -Dto=<toSiteId>
-Dfrom=<fromSiteId> -Dto=<toSiteId> -Daction=sites -Drestart=<restartAtSiteId>
-Dfrom=<fromSiteId> -Dto=<toSiteId> -Daction=groups -Drestart=<restartAtSiteId>
30639: Minor changes to commented out command line code for SiteServiceTestHuge
30643: Round of improvements to MySites dashlet relating to huge DB testing:
- 10,000 site database, user is a member of ~2000 sites
- Improvements to site.lib.ftl and related SiteService methods
- To return MySites dashlet for the user, order of magnitude improvement from 7562ms to 618ms in the profiler (now ~350ms in the browser)
30644: Fixed performance regression - too much opening and closing of the delta reader and writer
30661: More reader opening / closing
30668: Performance improvements to Site Finder and My Sites in user profile page.
- faster to bring back lists and site memberships (used by the Site Finder)
- related further improvements to APIs used by this and My Sites on dashboard
30713: Configuration for MySites dashlet maximum list size
30725: Merged V3.4-BUG-FIX to DEV/ALAN/SITE_PERF
30708: ALF-10040: Added missing ReferenceCountingReadOnlyIndexReaderFactory wrapper to IndexInfo.getMainIndexReferenceCountingReadOnlyIndexReader() to make it consistent with IndexInfo.getMainIndexReferenceCountingReadOnlyIndexReader(String, Set<String>, boolean) and allow SingleFieldSelectors to make it through from LeafScorer to the path caches! Affects ALL Lucene queries that run OUTSIDE of a transaction.
30729: Use getAuthoritiesForUser rather than getContainingAuthorities if possible.
SiteServiceTestHuge: command line version
30733: Performance improves to user dashboard relating to User Calendar
- converted web-tier calendar dashlet to Ajax client-side rendering - faster user experience and also less load on the web-tier
- improvements to query from Andy
- maximum sites/list size to query now configurable (default 100 instead of previously 1000)
30743: Restore site CRUD performance from cold caches
- Introduced NodeService.getAllRootNodes(), returning all nodes in a store with the root aspect, backed by a transactional cache and invalidated at key points
- Means indexing doesn't have to load all parent nodes just to check for 'fake roots'
- Site CRUD performance now back to sub-second with 60,000 nodes
30747: Improvement to previous checkin - prevent cross cluster invalidation of every store root when a single store drops out of the cache
30748: User dashboard finally loading within seconds with 60,000 sites, 60 groups, 100 users (thanks mostly to Kev's UI changes)
- post-process IBatis mapped statements with MySQL dialect to apply fetchSize=Integer.MIN_VALUE to all _Limited statements
- Means we can stream first 10,000 site groups without the MySQL JDBC driver reading all 240,000 into memory
- New NodeService getChildAssocs method with a maxResults argument (makes use of the above)
- Perfected getContainingAuthoritiesInZone implementation, adding a cutoff parameter, allowing only the first 1000 site memberships to be returned quickly and caches to be warmed for ACL evaluations
- New cache of first 10,000 groups in APP.SHARE zone
- Cache sizes tuned for 60,000 site scenario
- Site service warms caches on bootstrap
- PreferencesService applies ASPECT_IGNORE_INHERITED_RULES to person node to prevent the rule service trying to crawl the group hierarchy on a preference save
- WorkflowServiceImpl.getPooledTasks only looks in APP.DEFAULT zone (thus avoiding site group noise)
30749: Fix compilation errors
30761: Minor change to SiteServiceTestHuge
30762: Derek code review: Reworked fetchSize specification for select_ChildAssocsOfParent_Limited statement for MySQL
- Now fetchSize stated explicitly in a MySQL specific config file resolved by the HierarchicalResourceLoader
- No need for any Java-based post processing
30763: Build fix: don't add a user into its own authorities (until specifically asked to)
30767: Build fix
- IBatis / MySQL needs a streaming result statement to be run in an isolation transaction (because it doesn't release PreparedStatements until the end)
30771: Backed out previous change which was fundamentally flawed
- Resolved underlying problem which was that the select_ChildAssocsOfParent_Limited SQL string needs to be unique in order to not cause confusion in the prepared statement cache
30772: Backed out previous change which was fundamentally flawed
- Resolved underlying problem which was that the select_ChildAssocsOfParent_Limited SQL string needs to be unique in order to not cause confusion in the prepared statement cache
git-svn-id: https://svn.alfresco.com/repos/alfresco-enterprise/alfresco/HEAD/root@30797 c4b6b30b-aa2e-2d43-bbcb-ca4b014f7261