Query Troubleshooting

How to use the Couchbase Lite Query API’s explain() method to examine a query

Query Explain

Using

Query’s explain() method can provide useful insight when you are trying to diagnose query performance issues and-or optimize queries. To examine how your query is working, either embed the call inside your app (see Example 1), or use it interactively within a cblite shell (see Example 2).

Example 1. Using Query Explain in App

val query = QueryBuilder
    .select(SelectResult.all())
    .from(DataSource.collection(collection))
    .where(Expression.property("type").equalTo(Expression.string("university")))
    .groupBy(Expression.property("country"))
    .orderBy(Ordering.property("name").descending()) 

println(query.explain())

1. Construct your query as normal
2. Call the query’s explain method; all output is sent to the application’s console log.

Example 2. Using Query Explain in cblite

cblite <your-database-name>.cblite2 

(cblite) select --explain domains group by country order by country, name 

(cblite) query --explain {"GROUP_BY":[[".country"]],"ORDER_BY":[[".country"],[".name"]],"WHAT":[[".domains"]]} 

1. Within a terminal session open your database with cblite and enter your query
2. Here the query is entered as a N1QL-query using select
3. Here the query is entered as a JSON-string using query

Output

The output from explain() remains the same whether invoked by an app, or cblite—see Example 3 for an example of how it looks.

Example 3. Query.explain() Output

SELECT fl_result(fl_value(_doc.body, 'domains')) FROM kv_default AS _doc WHERE (_doc.flags & 1 = 0) GROUP BY fl_value(_doc.body, 'country') ORDER BY fl_value(_doc.body, 'country'), fl_value(_doc.body, 'name')

7|0|0| SCAN TABLE kv_default AS _doc
12|0|0| USE TEMP B-TREE FOR GROUP BY
52|0|0| USE TEMP B-TREE FOR ORDER BY

{"GROUP_BY":[[".country"]],"ORDER_BY":[[".country"],[".name"]],"WHAT":[[".domains"]]}

This output (Example 3) comprises three main elements:

1. The translated SQL-query, which is not necessarily useful, being aimed more at Couchbase support and-or engineering teams.
2. The SQLite query plan, which gives a high-level view of how the SQL query will be implemented. You can use this to identify potential issues and so optimize problematic queries.
3. The query in JSON-string format, which you can copy-and-paste directly into the cblite tool.

The Query Plan

Format

The query plan section of the output displays a tabular form of the translated query’s execution plan. It primarily shows how the data will be retrieved and, where appropriate, how it will be sorted for navigation and-or presentation purposes. For more on SQLite’s Explain Query Plan — see SQLite Explain Query Plan.

Example 4. A Query Plan

7|0|0| SCAN TABLE kv_default AS _doc
12|0|0| USE TEMP B-TREE FOR GROUP BY
52|0|0| USE TEMP B-TREE FOR ORDER BY

1. Retrieval method — This line shows the retrieval method being used for the query; here a sequential read of the database. Something you may well be looking to optimize — see Retrieval Method for more.
2. Grouping method — This line shows that the Group By clause used in the query requires the data to be sorted and that a b-tree will be used for temporary storage — see Order and Group.
3. Ordering method — This line shows that the Order By clause used in the query requires the data to be sorted and that a b-tree will be used for temporary storage — see Order and Group.

Retrieval Method

The query optimizer will attempt to retrieve the requested data items as efficiently as possible, which generally will be by using one or more of the available indexes. The retrieval method shows the approach decided upon by the optimizer — see Table 1.

Table 1. Retrieval methods

Retrieval Method	Description
Search	Here the query is able to access the required data directly using keys into the index. Queries using the Search mode are the fastest.
Scan Index	Here the query is able to retrieve the data by scanning all or part-of the index (for example when seeking to match values within a range). This type of query is slower than search, but at least benefits from the compact and ordered form of the index.
Scan Table	Here the query must scan the database table(s) to retrieve the required data. It is the slowest of these methods and will benefit most from some form of optimization.

When looking to optimize a query’s retrieval method, consider whether:

• Providing an additional index makes sense
• You could use an existing index — perhaps by restructuring the query to minimize wildcard use, or the reliance on functions that modify the query’s interpretation of index keys (for example, lower())
• You could reduce the data set being requested to minimize the query’s footprint on the database

Order and Group

The Use temp b-tree for lines in the example indicate that the query requires sorting to cater for grouping and then sorting again to present the output results. Minimizing, if not eliminating, this ordering and re-ordering will obviously reduce the amount of time taken to process your query.

Ask "is the grouping and-or ordering absolutely necessary?": if it isn’t, drop it or modify it to minimize its impact.

Queries and Indexes

Querying documents using a pre-existing database index is much faster because an index narrows down the set of documents to examine.

When planning the indexes you need for your database, remember that while indexes make queries faster, they may also:

• Make writes slightly slower, because each index must be updated whenever a document is updated
• Make your Couchbase Lite database slightly larger.

Too many indexes may hurt performance. Optimal performance depends on designing and creating the right indexes to go along with your queries.

Constraints

Couchbase Lite does not currently support partial value indexes; indexes with non-property expressions. You should only index with properties that you plan to use in the query.

The query optimizer converts your query into a parse tree that groups zero or more and-connected clauses together (as dictated by your where conditionals) for effective query engine processing.

Ideally a query will be able to satisfy its requirements entirely by either directly accessing the index or searching sequential index rows. Less good is if the query must scan the whole index; although the compact nature of most indexes means this is still much faster than the alternative of scanning the entire database with no help from the indexes at all.

Searches that begin with or rely upon an inequality with the primary key are inherently less effective than those using a primary key equality.

Working with the Query Optimizer

You may have noticed that sometimes a query runs faster on a second run, or after re-opening the database, or after deleting and recreating an index. This typically happens when SQL Query Optimizer has gathered sufficient stats to recognize a means of optimizing a suboptimal query.

If only those stats were available from the start. In fact, they are gathered after certain events, such as:

• Following index creation
• On a database close
• When running a database compact

So, if your analysis of the Query Explain output indicates a suboptimal query and your rewrites fail to sufficiently optimize it, consider compacting the database. Then re-generate the Query Explain and note any improvements in optimization. They may not, in themselves, resolve the issue entirely; but they can provide a useful guide toward further optimizing changes you could make.

Wildcard and Like-based Queries

Like-based searches can use the index(es) only if:

• The search-string doesn’t start with a wildcard
• The primary search expression uses a property that is an indexed key
• The search-string is a constant known at run time (that is, not a value derived during processing of the query)

To illustrate this we can use a modified query from the Mobile Travel Sample application; replacing a simple equality test with a LIKE.

In Example 5 we use a wildcard prefix and suffix. You can see that the query plan decides on a retrieval method of Scan Table.

Tip

For more on indexes — see Indexing

Example 5. Like with Wildcard Prefix

val query = QueryBuilder
    .select(SelectResult.all())
    .from(DataSource.collection(collection))
    .where(
        Expression.property("type").like(Expression.string("%hotel%"))
            .and(Expression.property("name").like(Expression.string("%royal%")))
    )
println(query.explain())

The indexed property, type, cannot use its index because of the wildcard prefix.

Resulting Query Plan

2|0|0| SCAN TABLE kv_default AS _doc

By contrast, by removing the wildcard prefix % (in Example 6), we see that the query plan’s retrieval method changes to become an index search. Where practical, simple changes like this can make significant differences in query performance.

Example 6. Like with No Wildcard-prefix

val query = QueryBuilder
    .select(SelectResult.all())
    .from(DataSource.collection(collection))
    .where(
        Expression.property("type").like(Expression.string("hotel%"))
            .and(Expression.property("name").like(Expression.string("%royal%")))
    )
println(query.explain())

Simply removing the wildcard prefix enables the query optimizer to access the typeIndex, which results in a more efficient search.

Resulting Query Plan

3|0|0| SEARCH TABLE kv_default AS _doc USING INDEX typeIndex (<expr>>? AND <expr><?)

Use Functions Wisely

Functions are a very useful tool in building queries, but be aware that they can impact whether the query-optimizer is able to use your index(es).

For example, you can observe a similar situation to that shown in Wildcard and Like-based Queries when using the lower() function on an indexed property.

Query

val query = QueryBuilder
    .select(SelectResult.all())
    .from(DataSource.collection(collection))
    .where(Function.lower(Expression.property("type")).equalTo(Expression.string("hotel")))
println(query.explain())

Here we use the lower() function in the Where expression

Query Plan

2|0|0| SCAN TABLE kv_default AS _doc

But removing the lower() function, changes things:

Query

val query = QueryBuilder
    .select(SelectResult.all())
    .from(DataSource.collection(collection))
    .where(Expression.property("type").equalTo(Expression.string("hotel"))) 
println(query.explain())

Here we have removed lower() from the Where expression

Query Plan

3|0|0| SEARCH TABLE kv_default AS _doc USING INDEX typeIndex (<expr>=?)

Knowing this, you can consider how you create the index; for example, using lower() when you create the index and then always using lowercase comparisons.

Optimization Considerations

Try to minimize the amount of data retrieved. Reduce it down to the few properties you really do need to achieve the required result.

Consider fetching details lazily. You could break complex queries into components. Returning just the doc-ids, then process the array of doc-ids using either the Document API or a query that uses the array of doc-ids to return information.

Consider using paging to minimize the data returned when the number of results returned is expected to be high. Getting the whole lot at once will be slow and resource intensive. Plus does anyone want to access them all in one go? Instead, retrieve batches of information at a time, perhaps using the LIMIT/OFFSET feature to set a starting point for each subsequent batch. Although, note that using query offsets becomes increasingly less effective as the overhead of skipping a growing number of rows each time increases. You can work around this, by instead using ranges of search-key values. If the last search-key value of batch one was 'x' then that could become the starting point for your next batch and-so-on.

Optimize document size in design. Smaller docs load more quickly. Break your data into logical linked units.

Consider Using Full Text Search instead of complex like or regex patterns — see Full Text Search.