On a few projects in 2014, the issue of Endeca's performance came up. Specifically, applications were seeing a large number of queries and were also generating large response sizes from Endeca. These queries were not being generated by the Assembler API, but were one-off queries created to bring back other information from Endeca.
This article will give some tips on how to optimize those queries.
A response from Endeca can consist of a number of different pieces of data:
In an Assembler based application, the overall flow for a standard page being rendered would be this:
First, a super-lightweight query is executed that only returns Supplemental objects representing information from Experience Manager. Depending on that meta-data, one or more subsequent queries is executed that will return:
Thus a standard basic page would generate about two Endeca queries to be rendered. There are some cartridges that generate more:
In addition, the Assembler API is usually very good about only bringing back the exact information it needs. This means that it will only "open up" the dimensions being requested, and will return back the attributes on the products specified in the ResultsList configuration.
When it comes to making standalone queries to Endeca, you need to understand the information above so as to NOT bring back any more data than necessary.
The easiest way to scan an application for one-off Endeca queries is to do a search for "ENEQuery" or "ENEQueryResults" in you .java classes. In addition, searching for "InvokeAssembler" in .jsp's. You can also search for UrlENEQuery.
If you find many instances of those, each and every query involved should be assessed using the information laid out below.
In a standard commerce application, it's not uncommon for one record (representing a product or SKU) to have 80 or 100 or more properties. These can be things like product ID, UPC, short descriptions, size/color/widths, prices, image URLs, etc.
If you are not careful, it's very easy to return all 80 or 100 or whatever property values with a standalone query.
To look at what comes back by default, you can look at the orange JSP reference application (typically located at http://localhost:8006/endeca_jspref, with a host of localhost and port of 15000).
The list of properties to be returned can be controlled by using ENEQuery.setSelection(). This requires that you to specify every single property (and dimension) to be returned. It is case-sensitive.
By default, a query will return 10 records. To limit this, you can use ENEQuery.setNavNumERecs.
In the .reqlog, if you see &nbins=10, that means that someone didn't set this value specifically and is probably using the default.
At the same time, you shouldn't set this value to be too large. If you find yourself setting this to 50 or 100, you might be doing something wrong.
A supplemental object is the meta-data about a landing page or content slot. If you use the orange reference application, at the top you'll see one or more "Unknown Merch Style". Scrolling to the bottom of page, you'll see a series of "Matching Supplemental Objects".
What's the big deal about these? Well, these can actually get somewhat large in size (for instance, if you have cartridges that allow merchandisers to copy/paste raw HTML). Also, the only real time they need to come back is when doing Assembler queries; not one-off queries.
There's no flag for turning supplemental objects on/off. However, you can add a merch rule filter that will have the effect of turning them off. (This is what a hotfix for 10.2 and what 11.x do by default. If you look in the .reqlog, you'll see &merchrulefilter=endeca.internal.nonexistent in some of the queries).
This can be turned on by using ENEQuery.setNavMerchRuleFilter(). Basically any nonsense string in here will have the correct effect. This would also be a good place to put a message in for logging purposes. Something like ENEQuery.setNavMerchRuleFilter("topNavigationQuery").
In the .reqlog, you should see &merchrulefilter .
If you look at the orange reference app, you'll see that the dimensions on the left side are "closed" up. If you click one, the page will refresh and now that dimension will be "opened" up.
If you would like to open up all dimensions, you can use ENEQuery.setNavAllRefinements(true).
However, this can be potentially very expensive. With no dimensions being returned, the MDEX Engine doesn't have to compute the refinement counts (aka, "How many records are there for Brand=XYZ"?) Also, this can inflate the response size greatly, especially for big flag dimensions.
Instead, you should specify which particular dimensions you want to return. Unfortunately, you need to specify the ID of the dimension, not the name.
If you know the ID of the dimensions you care about, you can use UrlENEQuery.setNe() and pass in a string like "123+234+532".
Looking through the .reqlog, if you see &allgroups=1, that means somewhere someone has setNavAllRefinements(true).
Let's say you're on a product details page. If you know the ID of the product, you have two choices: You can do a keyword search on the ID field passing in the string of the value. Or you can construct a record filter. A record filter is usually faster and cleaner. (There's no reason to fill your logs with searches that customers didn't type in).
ENEQuery.setNavRecordFilter() is the method. An example might be: query.setNavRecordFilter("AND(product.id:2342342))".
A little-used feature is the ENEQuery.setQueryInfo() method. This lets you stuff any number of key/value pairs that get sent to the MDEX Engine, ignored, but written out to the .reqlog file. This can be useful for adding things like session ID, debug information, etc.
For our case, what might be good is to write out why this query is being executed. "pdpBreadcrumbs" "typeahead" , etc.
This way, if there are slow or big queries found during performance testing, it will help track them down and help distinguish between real Assembler queries and your one-off queries.
These messages will show up in the .reqlog as &log=
ENEQuery.setNavERecsPerAggrERec() allows you to specify how many records are returned per aggregate record. For example, say you are a clothing website. You probably index by SKU (which would represent a single Size/Color combination for a product). When doing query to Endeca, instead of returning info at a SKU level, you would aggregate things by a rollup key using ENEQuery.setNavRollupKey().
setNavERecsPerAggrERec() allows to you bring back 0, 1 or all SKUs within a product. You should do everything possible to NOT set it to the value of "2", which is all.
(As a point of reference, ENEQuery has 3 static values representing those numbers. ZERO_ERECS_PER_AGGR, ONE_EREC_PER_AGGR, ALL_ERECS_PER_AGGR).
In the .reqlog, if you see &allbins=2, then that means someone setNavERecsPerAggrERec(ALL_ERECS_PER_AGGR).
Now, this might make things complicated for you. For instance, on one site on the search results page, they wanted to display the color swatches from each different SKU. By setting it to bring back all, they were able to iterate across all of the SKUs in the product to generate that list.
Instead, things were changed so that each SKU was tagged with information about all of the other SKUs. This allowed us to change this from 2 to 1. Response sizes went from 10 megs to 100kb.
For some commerce sites, they might set products to activate during the day ("Starting at 1pm EST, this product should show up, but before that, it shouldn't").
One way to do this would be to tag all products with a start date and end date. And then with each query to Endeca, pass along a range filter for the dates.
The problem, however, can be that the MDEX Engine does some internal caching based on these values. If the date value you specify is too granular, then the MDEX won't work as fast as it could. So don't specify a timestamp down to the second or millisecond. Try and do timestamps for the hour, or at least chunks of minutes (like 20 or 30 minutes) to ensure that some cache hits occur.
Range filters can be set by using setNavRangeFilters().
In the .reqlog, you can look for &pred . A CRS example might look like: pred=product.endDate%7cGTEQ+1.4163552E12&pred=product.startDate%7cLTEQ+1.4163552E12
This is a little-used feature, so it's not something you'd come across very often. Key properties return meta-data about the definitions of properties and dimensions themselves. This can be turned on using ENEQuery.setNavKeyProperties(ENEQuery.KEY_PROPS_ALL).
This can greatly inflate the response size of a query from Endeca.
If you do need this for some reason, you should only need to execute the query once, and then cache the results from it.
This can be found in the .reqlog as &keyprops=all
Careful readers might notice that CRS breaks some of the rules above. In particular:
As an interesting point of reference, the world's worst Endeca query would: