ZK Huge Grouping Model"

From Documentation
Line 409: Line 409:
 
</source>
 
</source>
  
Nothing remarkable here either, initialize the model, and handle paging events (updateOffsetIndex(x) in the delegating page model), the rest is done in PagingGroupsModel.
+
Nothing remarkable here either
 +
* initialize the model
 +
* handle paging events (updateOffsetIndex(x) in the delegating page model)
  
Running the application
+
The rest is done in PagingGroupsModel.
 +
 
 +
==Running the application==
  
 
     mvn jetty:run
 
     mvn jetty:run
 
     localhost:8080/hugeGrouping/accesslog_mvvm.zul
 
     localhost:8080/hugeGrouping/accesslog_mvvm.zul
  
The application will be very responsive, where ever you navigate, opening closing groups. Looking at the console log, you'll get some information about Cache hits and Cache MISSES.
+
The application will be very responsive, where ever you navigate, opening closing groups. Looking at the console log, you'll get some information about Cache hits and Cache MISSES, notice the high number of cache hits after a randomly navigating to a high page.
 +
 
 +
The initial setup of the AccessLogDao (mock) is 40 million groups with 8 to 88 children randomly distributed, so we roughly end up with 2 BILLION rows.
 +
 
 +
The memory consumption will remain stable once the Dao is initialized (these 160MB - mentioned above - would usually be on the DB and not in memory) when navigating the list.
 +
Once the cache of the binary grouping position search is hot, the number of DB calls is low (usually around 1-4 less frequently bigger) per page increment / decrement, after randomly accessing any arbitrary page (despite some special cases).
 +
 
 +
And the data returned from the DB is very little:
  
The initial setup of the AccessLogDao is 400 million groups with 8 to 88 children randomly distributed, so we roughly end up with 2 BILLION rows.
+
* loading 1 or 2 (or several very small) group children chunks
 +
* loading 0 or 1 new chunks of group informations
 +
* counting the ranges that earn Cache MISSES during the binary search for the next page's startIndex (could be avoided completely with a more intelligent search algorithm, being aware of single page increments)
  
 
=Appendix=
 
=Appendix=

Revision as of 10:16, 2 August 2013

ZK Huge Grouping Model

Author
Robert Wenzel, Engineer, Potix Corporation
Date
August, 2013
Version
ZK 6.5 (or later)

WarningTriangle-32x32.png This page is under construction, so we cannot guarantee the accuracy of the content!

Introduction

bla bla you have some big data... how to display

article already handles display big data in a flat list http://books.zkoss.org/wiki/Small_Talks/2009/July/Handling_huge_data_using_ZK

based on the concepts there (paging at DB level, separate paging control from grid) how to do grouping...

The Challenges

Grouping is a powerful feature in ZK but also adds an extra layer of complexity.

Paging and grouping challenges

1. groups can be open or closed (initially and interactively)

-> the total count and the number of pages changes, when opening/closing nodes... (needs efficient counting, and state keeping)

2. groups can have arbitrary number of children

-> random access to a specific page ... how to know the current group and position inside the group for that page
--> implement a feasible search

3. minimize DB operations (accumulating network/DB latency)

-> caching vs. memory consumption

1. + 2. + 3. !!!! combining all three in an efficient, memory preserving way

Limitations

  • needs to store the open/close state in memory (humans are limited, so one is unlikely to toggle 100+ groups, just wanted to mention this)
  • only works on static data - when data changes, the model needs to be reinitialized and caches need to be cleared
  • as not all rows are present in memory, you cannot send UI events to components currently not on the screen
  • some dummy components are required to make the <grid> accept the limited view over the data

One might argue whether it is useful to display such a long list in this way (especially without pre-filtering the results).

We are here to show that it is possible, maxing out the limits ... ZK GroupsModel supports indexes of int-datatype so we will go up to somewhere around Integer.MAX_VALUE (~ 2.000.000.000 records)

Implementation

Accessing the data

not use DB, just deterministic random value... caching the group sizes in memory e.g. 400.000.000 of groups child counts stored in an int[] are still 160 MB of memory

Data Record and Dao

package org.zkoss.grouping.dao;

import java.util.Date;

public class AccessDataRecord {
	private String ipAddress;
	private String browser;
	private long contentLength;
	private String country;
	private Date accessTime;
	private String url;

//... Constructor + getters
package org.zkoss.grouping.dao;

public class AccessLogDao {

	public int getGroupCount() {
		//your DB query to count the groups
	}

	public List<GroupInfo<String, String>> getGroupInfos(int startIndex, int pageSize) {
		//your DB query and mapping to create GroupInfos for this page of groups
	}

	public List<AccessDataRecord> getChildInfos(int groupIndex, int startIndex, int pageSize) {
		//your DB query and load this page of children in the group
	}

	/**
	 * This method will count a lot and return a constant value as long as data does not change, 
	 * so caching would be desirable
	 * @return
	 */
	public int getTotalChildCount() {
		//your DB impl to calc the total number of children
	}

	/**
	 * This method will be called quite often with recurring parameters, 
	 * so caching (of most frequently used params) would be desirable !!!
	 * @return the childrenCount between 2 groups... (including groupIndexFrom and excluding groupIndexTo)
	 */
	public int getChildCountBetween(int groupIndexFrom, int groupIndexTo) {
		//your DB impl to calc the number of children between 2 groups
	}

The Paging Model

AccessDataGroupsModel our specialized PagingGroupsModel<D, H, F> is using AccessDataRecord as "data"(D) and String as "head"(H) and "foot"(F) implementation. This class implements the 4 model methods and a GroupingPositionSearch strategy - all delegating to AccessLogDao from above, to keep the workload at DB-Level.

package org.zkoss.grouping;

import java.util.List;

import org.zkoss.grouping.dao.AccessLogDao;
import org.zkoss.grouping.dao.AccessDataRecord;
import org.zkoss.grouping.model.GroupInfo;
import org.zkoss.grouping.model.PagingGroupsModel;
import org.zkoss.grouping.model.search.BinaryGroupingPositionSearch;
import org.zkoss.grouping.model.search.GroupingPositionSearch;

class AccessDataGroupsModel extends
		PagingGroupsModel<AccessDataRecord, String, String> {

	private AccessLogDao groupsDao;

	public AccessDataGroupsModel(AccessLogDao groupsDao, int pageSize, boolean initialOpen, boolean hasGroupfoot) {
		super(pageSize, initialOpen, hasGroupfoot);
		this.groupsDao = groupsDao;
		setPositionSearch(binarySearch());
	}

	@Override
	protected int loadGroupCount() {
		return groupsDao.getGroupCount();
	}

	@Override
	protected List<GroupInfo<String, String>> loadGroupPage(int startIndex, int pageSize) {
		return groupsDao.getGroupInfos(startIndex, pageSize);
	}

	@Override
	protected List<AccessDataRecord> loadChildrenPage(int groupIndex, int startIndex, int pageSize) {
		return groupsDao.getChildInfos(groupIndex, startIndex, pageSize);
	}

	@Override
	protected int getTotalChildCount() {
		return groupsDao.getTotalChildCount();
	}

	private GroupingPositionSearch binarySearch() {
		return new BinaryGroupingPositionSearch(this, 4096) {
			@Override
			protected int getChildCountBetween(int groupIndexFrom, int groupIndexTo) {
				return groupsDao.getChildCountBetween(groupIndexFrom, groupIndexTo);
			}
		};
	}

}

As we see the methods are focusing on counting, and retrieving data in page sized chunks. Nothing magical here, however our DB is doing the heavy work (selecting, grouping, counting, sorting) ... that's what it is optimized for, and with huge data we'll not attempt to compete with our DB, we just want as little data as possible.

Page/Caching, Counting & Open/Close State keeping

Open/Close State keeping

The PagingGroupsModel will keep track of the toggled state only for the changed groups (together with the childCount of that group). If the Model is initialized with INITIALLY_OPEN, then only the closed nodes will be tracked, and vice versa. Assuming a human will not take the time to toggle an outrageous number of groups, this Map (groupIndex => childcount) will stay relatively small and should not compromise our performance and memory expectations. Additionally the total number of toggled groups is updated on every open/close interaction.

This information is used to adjust the number of total children (for the UI paging calculations) or the number of children between 2 groups (for Position search) - as the DB won't keep that UI state.

see:

  • PagingGroupsModel#getCurrentRowCount()
  • PagingGroupsModel#getToggledCountBetween(int groupIndexFrom, int groupIndexTo)
  • PagingGroupsModel#toggleGroup(int groupIndex, boolean open)

Page/Caching

The GroupsModel interface offers methods to retrieve single Groups, and Children... Whenever one of these methods is called the PagingGroupsModel will check if that position is cached for the current page, and reload a page sized chunk from the DB if not, and cache that chunk both at group- and child-level.

One problem here is, that the chunk retrieved from the DB will not match the page boundaries in the UI.

e.g.

    1         2         3         4                
    |.........|.........|.........|.........           -> UI pages (page size 10)

    0.....1.........|......2...3......4....5           -> Groups | Chunks Group 0 open
    01.........|......2...3......4....5.....           -> Groups | Chunks Group 0 closed

Page size vs. Group size: While the UI-Pages have a constant size the groups may have a variable size, and their offset may shift because of previously opened/closed nodes, the offset of the chunks loaded from the DB would vary, whenever nodes are toggled.

As we want to benefit from caching at all possible levels, keeping the method/query parameters constant will reduce our required cache size at Dao method level or in DB and will avoid duplicate load of overlapping chunks (with multiple users having different UI state).

So in the scenario above whenever e.g. Element 3 of Group 1 needs to be loaded the PagingGroupsModel will load the chunk "Group1 Child0-9" calling this.loadChildrenPage(1, 0, 10). Perfect for caching, as it would never call e.g. this.loadChildrenPage(1, 3, 10).

see:

  • AccessDataGroupsModel#loadGroupPage(int startIndex, int pageSize)
  • AccessDataGroupsModel#loadChildrenPage(int groupIndex, int startIndex, int pageSize)

The same strategy applies to caching the Groups. PagingGroupsModel will also load chunks of Groups with fixed intervals. In case groups are displayed with a foot, the groups cache size can be halved, as the group foot will consume one row too.

cachedGroups = Collections.synchronizedMap(new LRUMap(pageSize * (hasGroupFoot ? 1 : 2))); //cache at least 2 pages ...
cachedChildren = Collections.synchronizedMap(new LRUMap(pageSize * 4)); //cache enough for at least a few pages

The cache size is reasonably small and the LRUMap will make sure it will not grow beyond its capacity throwing out the oldest Groups/Children, when overflowing.

The disadvantage of this approach is, that in most cases at least two chunks need to be loaded to display a single page, however when caching a few of these chunks in the model, you'll have one of the two already loaded for the previous page no matter which direction you navigate.

I think the better cachability outrules the slight memory overhead. At a page size of 50, we might have 49 records too much in memory (which will likely be reused on the next page, when the user navigates, by one Page). But reduce the permutations of method/query parameters by 98% which highly improves the chance of a cache hit at method or DB level. I would even keep the cache bigger, so that back navigation by few pages does not cause a reload.

Also very small groups will result in more DB round trips, but hey ... we are talking about BIG data :)

Random Access Paging

When just navigating page by page, this would already be sufficient... but what if you want to navigate to page no 234.455.234 :-o

This is no problem when Groups are closed by default, as each group will only consist of a Head (or Head and Foot).

If pages are open by default you suddenly encounter a question you might not even have thought about before, when doing pagination in a flat list without the grouping level - plus some groups open/closed.

Which group- / child-index does page X start with??

Of course the brute force way, just iterating over all the groups accumulating their size until the (page * pageSize) is reached is not feasible. It will either cause many DB operations, or one big traversal over a huge result set of Group counts (Keep in mind, we don't know how many groupCounts we need to accumulate until we reach the page X).

I am not a DB expert, so maybe there is a solution for this problem already present at DB level (and we'd have to adjust the result by the number of closed nodes the DB does not know about) and I am not a scientist so there is maybe a more sophisticated perfectly balanced search algorithm - Please comment if you know :)

So here is my approach which will keep the required DB operations at a low level, and again provide high chance for cache hits, especially for the most expensive often recurring count operations.

A binary search:

1. Bisect the number of groups into Ranges and have the DB count the rows for that range

e.g. (I am sure the DB is faster than you in java code (assuming efficient indexing) ;))
SELECT count(*) WHERE group_key >= x AND group_key < y

2. Check whether the result (adjusted by group-heads(, -feet) and the toggled groups) is bigger or smaller.

3. Define the next smaller range always halving the intervals.

Repeat 1-3 that until you find the group.

  • For 1024 groups you'll not have to repeat that more than 9 times until you find the group we are interested in.
  • For 1024 * 1024 (about a million) it is only 19 times
  • For 1024 * 1024 * 1024 (about a billion) it is only 29 times

So you see this scales well 2x the data, requires only 1 additional DB round trip.

When starting with 1024 Groups a possible sequence of ranges might be

  1. 0 < 512 (WHERE group_key >= 0 AND group_key < 512)
  2. 0 < 256 ...
  3. 128 < 192
  4. 128 < 160
  5. 144 < 160
  6. 152 < 160
  7. 156 < 160
  8. 156 < 158
  9. 157 < 158 :D the page start must be in in group #157

The icing on the cake is: The most counting intensive queries will only have little variance in their query parameters, so they will be most likely in the cache already, and the memory consumption for one cache entry is very low... 3 integers (range start and end as cache key, and another integer for the count) -> Caching a few hundred(or thousand) of those results does not weigh too much, in contrast to caching each group count in memory (that's up to the DB)

1st query is always the same 2nd query has only 2 options 3rd query has 4 options ...

Still we have a penalty of several DB queries for a random access to a page, but after that the cache is hot, and single page navigation will take a similar path in the binary search, causing much fewer cache misses (except for some rare cases, where you cross the boundary of a bigger range, but still the penalty will never be as big as the first request). Additionally the ranges will be the same for concurrent users, offering the possibility to cache at a lower application level too.

Since the user will most likely begin at the first page, I added an alternative approach starting from group zero, like a bottom-up search, doubling the ranges in each iteration until the searchIndex is exceeded and then refine the result reusing the binary approach. This postpones the expensive counting operations to a time when a "random access occurs"... as I said I am not a Scientist, so I just defined a threshold for the searchIndex. For bigger search indexes the top-to-bottom search is used.

For the case someone has a better strategy, I extracted the GroupingPositionSearch interface for everyone to implement themselves (maybe a pure DB stored procedure).

See:

  • org.zkoss.grouping.model.search.GroupingPositionSearch
  • org.zkoss.grouping.model.search.LinearGroupingPositionSearch
an implementation for fun... to compare the performance difference
  • org.zkoss.grouping.model.search.BinaryGroupingPositionSearch
all one needs is to implement the method getChildCountBetween() and query the DB

Enough pseudo theoretical talk ... let's use this!

Using the Model in the UI

We have seen how to access the data, and how the caching will hopefully have some positive effects. All that is remaining is the ViewModel, and the View.

accesslog_mvvm.zul

	<window border="normal" height="810px" width="1200px" 
			apply="org.zkoss.bind.BindComposer" viewModel="@id('vm') 
			@init('org.zkoss.grouping.AccessDataGroupingViewModel')">
		
		<paging pageSize="@load(vm.pageSize)" 
			activePage="@bind(vm.activePage)" 
			totalSize="@load(vm.totalSize)"
			onPaging="@command('changePage', activePage=event.activePage)"
			detailed="true" />
			
		<grid model="@load(vm.currentPageModel)" >
			<columns>
				<column label="ip" ></column>
				<column label="url" ></column>
				<column label="time" ></column>
				<column label="country" ></column>
				<column label="browser" ></column>
				<column label="length" ></column>
			</columns>

			<template name="model:group">
				<group label="@init(each)" visible="@init(not(vm.isDummy(each)))" ></group>
			</template>
			
			<template name="model">
				<row>
					<label value="@init(each.ipAddress)"></label>
					<label value="@init(each.url)"></label>
					<label value="@init(each.accessTime)"></label>
					<label value="@init(each.country)"></label>
					<label value="@init(each.browser)"></label>
					<label value="@init(each.contentLength)"></label>
				</row>
			</template>
			<template name="model:groupfoot">
				<groupfoot visible="@init(not(vm.isDummy(each)))">
					<cell colspan="6" sclass="z-groupfoot-inner">
						<label value="@init(each)" sclass="z-groupfoot-cnt" />
					</cell> 
				</groupfoot>
			</template>
		</grid>
	</window>

Pretty basic Grouping Grid not using the paging mold, but a separate paging control, as we only send a delegating Model (model="@load(vm.currentPageModel)") to the <grid> component. This is to only present as little information to the grid component to avoid the component grid from caching data internally, and full control over what is loaded.

Because the Grid component expects groups to open and close in a consistent way, a dummy head and foot, need to be added if a Page does not start with a group head, or end with a foot. (this is slightly unpleasant, but what do we get in return) ... the biggest ZK grouping grid in history.

AccessDataGroupingViewModel

public class AccessDataGroupingViewModel {
	private static final String DUMMY_FOOT = new String();
	private static final String DUMMY_HEAD = new String();
	//TODO: inject
	private AccessLogDao groupsDao = AccessLogDao.getInstance();
	private PagingGroupsModel<AccessDataRecord, String, String> groupsModel;
	
	private int pageSize;
	private int activePage;
	
	private DelegatingSinglePageGroupsModel<AccessDataRecord, String, String> currentPageModel;
	
	@Init
	public void init() {
		pageSize = 20;
		activePage = 0;
		groupsModel = new AccessDataGroupsModel(groupsDao, pageSize, 
				PagingGroupsModel.INITIALLY_OPEN, PagingGroupsModel.GROUPFOOT_ON);
		currentPageModel = groupsModel.getSinglePageModel(DUMMY_HEAD, DUMMY_FOOT);
		refreshRows();
	}
	
	@Command("changePage")
	@NotifyChange("currentPageModel")
	public void onChangePage() {
		refreshRows();
	}
	
	private void refreshRows() {
		currentPageModel.updateOffsetIndex(activePage * pageSize);
	}
	
	public int getPageSize() {
		return pageSize;
	}
	
	public int getActivePage() {
		return activePage;
	}
	
	public void setActivePage(int activePage) {
		this.activePage = activePage;
	}
	
	public int getTotalSize() {
		return groupsModel.getCurrentRowCount();
	}

	public GroupsModel<AccessDataRecord, String, String> getCurrentPageModel() {
		return currentPageModel;
	}
	
	public boolean isDummy(String string) {
		return DUMMY_HEAD == string || DUMMY_FOOT == string ;
	}
}

Nothing remarkable here either

  • initialize the model
  • handle paging events (updateOffsetIndex(x) in the delegating page model)

The rest is done in PagingGroupsModel.

Running the application

   mvn jetty:run
   localhost:8080/hugeGrouping/accesslog_mvvm.zul

The application will be very responsive, where ever you navigate, opening closing groups. Looking at the console log, you'll get some information about Cache hits and Cache MISSES, notice the high number of cache hits after a randomly navigating to a high page.

The initial setup of the AccessLogDao (mock) is 40 million groups with 8 to 88 children randomly distributed, so we roughly end up with 2 BILLION rows.

The memory consumption will remain stable once the Dao is initialized (these 160MB - mentioned above - would usually be on the DB and not in memory) when navigating the list. Once the cache of the binary grouping position search is hot, the number of DB calls is low (usually around 1-4 less frequently bigger) per page increment / decrement, after randomly accessing any arbitrary page (despite some special cases).

And the data returned from the DB is very little:

  • loading 1 or 2 (or several very small) group children chunks
  • loading 0 or 1 new chunks of group informations
  • counting the ranges that earn Cache MISSES during the binary search for the next page's startIndex (could be avoided completely with a more intelligent search algorithm, being aware of single page increments)

Appendix

Download

selenium-IDE-example.zip

Comments



Copyright © Potix Corporation. This article is licensed under GNU Free Documentation License.