GraphConnect 2022 - Top 10 Cypher Tuning Tips & Tricks.pptx

© 2022 Neo4j, Inc. All rights reserved.
Top 10 Cypher
(Tuning) Tips & Tricks
Michael Hunger,
Senior Director, User
Innovation

3
Why should I care?
1 2
Efficiency Readability
3 4
Resources Network

5
Dataset - (All of) StackOverflow
https://demo.neo4jlabs.com:7473
username/password/database: stackoverflow
51M nodes
124M relationships

6
1. PROFILE

7
1. Profile
• PROFILE and EXPLAIN show you the query plan
• All Operators – Index lookups, Expand, Projection, …
• Cardinality (rows) - multiplicative
• DB-Hits (measure of cost)
• Avoid Eager for Updates
• Shows Memory Usage
• Best if you can stream through and shortcut (e.g. LIMIT)
• avoid aggregation / sorting / collect if you can

8
PROFILE
PROFILE
MATCH (t:Tag)<-[:TAGGED]-(q:Question)
RETURN t.name, count(*) AS c
ORDER BY c DESC LIMIT 10
// Cypher version: CYPHER 4.4,
// planner: COST,
// runtime: PIPELINED.
// 97726446 total db hits
// in 29387 ms

9
2. Properties

10
2. Properties
• Property reads are more expensive than they should be
• Avoid them if possible, or defer to the latest moment
• When you have narrowed down the data to what you want to return
• Elevate properties to labels or rel-types can speed up a lot
• E.g. boolean, status properties as labels
• Sub-Structure of rel-properties to Rel-Types (e.g. year of a datetime)
• Use indexes, properties can be read from the index (except for points)
which is much faster (need indicate type)

11
PROFILE
PROFILE
MATCH (t:Tag)<-[:TAGGED]-(q:Question)
WITH t, count(*) AS c
RETURN t.name, c
// 97526566 total db hits
// in 29831 ms.

12
3. Relationships

13
Relationships
• Avoid traversing relationships if you can, each is O(1) but all are O(n)
• Always provide rel-type and direction
• Don’t traverse across supernodes, check “against” them
• Use get-degree (size(pattern)) without end label to get fast reads
• Elevate properties to rel-type for speedups
• Use relationship-indexes for global lookups or
• Add source/target key/property to index for node-centric indexes

14
PROFILE
PROFILE
MATCH (t:Tag)
WITH t, size([p=(t)<-[:TAGGED]-()|p]) AS c
RETURN t.name, c
// 104911 total db hits in 195 ms

15
4. Indexes & Constraints

16
PROFILE
PROFILE
MATCH (t:Tag)
RETURN t.name, t.count AS c
// 157356 total db hits in 66 ms

17
PROFILE
PROFILE
MATCH (t:Tag)
// type triggers index usage
WHERE t.count > 0
// 32 total db hits in 1 ms.

18
Types & Uses of indexes and constraints
1 2
Lookup
Ranges
Type Scans
3 4
Sorting
Aggregation
Uniqueness
Existence

19
Indexes & Constraints
• show indexes
• show constraints
• create <type> index/constraint <name> IF NOT EXISTS ON (x) FOR
(prop)
• todo docs

20
Indexes & Constraints
• lookup
• lookup in list
• lookup range
• don't read from property store
• sorting / aggregation
• compound indexes (one range, rest equality)
◦ NOM type
• relationship(type) indexes
◦ global lookup by attribute / range
◦ simulate node-centric indexes index start/end-node id or property

21
neo4j.com/docs/cypher-manual/current/query-tuning/using/
5. Query Hints

22
Query Hints
• shouldn't be needed
• but planner is not smart enough (yet)
• use rarely, first try without
• USING INDEX [SEEK] - force index usage (also for relationships)
• USING JOIN ON - force hash join for two high-cardinality sides
◦ example with LDBC
• USING SCAN ON - force label scan if very selective label

23
INDEX Hint
PROFILE
MATCH (t:Tag)
WHERE t.count > 0
USING INDEX :Tag(count)
// 32 total db hits in 1 ms.

24
4. Var Length Queries

25
4. Var Length Queries
• Traverse arbitrary length paths / trees
• Limit length if possible
• Pruning Var-Length Expand ( on distinct end nodes )
• Shortest Path ( can also be used as condition )

26
Var Length Example
profile
MATCH (u:User {name:'meistermeier'})-
[:POSTED|ACCEPTED|ANSWERED*..5]-(u2:User)
RETURN count(u)
// 569.130 total db hits

27
Var Length Example
profile
MATCH (u:User {name:'meistermeier'})-
[:POSTED|ACCEPTED|ANSWERED*..5]-(u2:User)
RETURN count(distinct u)
// 217.347 total db hits

28
Triadic Closure Example
profile
MATCH (u:User {name:'meistermeier‘})
--()--(u2:User)
WHERE NOT (u)--(u2)
RETURN count(distinct u)
// 810 total db total db hits

29
5. Fun with Lists

30
5. Fun with Lists
• Like in Python
• [ elem IN list WHERE pred(elem) | expr(elem) ]
• Result is a list again
• Quantors
• all/single/none/any(x IN list WHERE pred(element))
• List into rows: UNWIND list AS elem
• Rows into list: collect(elem) AS list

31
6. Map Projections

32
5. Map Projections
• Idea borrowed from GraphQL
• Map expression - node/relationship/map as expression
• var { .* } // all properties
• var {.property1, .property2}
• var {value} // key = „value“
• var {name: expression} // any expression, incl. map/list
• collect map expression to get list of documents
• You can use map { .*, foo : null } to remove properties from a map
(e.g. for LOAD CSV)

33
5. Map Projections
-[:POSTED]->()-[:ANSWERED]->(q)-[:TAGGED]->(tag)
WITH q, collect(tag.name) as tags
RETURN q { .title, .answers, .views, tags, } as question

34
7. Pattern
Comprehensions

35
6. Pattern Comprehensions
• Idea borrowed from GraphQL
• Pattern comprehension = like list comprehension just for patterns
• Pattern (can introduce new local variables)
• Filter
• Expression
• [ (pattern)-[of:SOME]->(nodes) WHERE filter | expression ]
• Result is list of the expression
• Expressions can be pattern comprehensions again or map expressions
• Not well planned (indexes etc)
• No sorting / pagination / pushdown

36
6. Pattern Comprehensions
RETURN [(u)-[:POSTED]->()-[:ANSWERED]->(q) |
q { .title, .answers, .views }] as questions

37
8. Subqueries

38
6. Subqueries
• Added in 4.0/4.1
• CALL { WITH var MATCH … RETURN var2 } …
• Existential subqueries EXISTS {}
• Can not shadow variables in RETURN
• Dependent or independent
• Still participate in cardinality but can change it
• Properly planned, unlike pattern comprehensions -> can rewrite

39
9. Batched Updates

40
7. Batched Updates
• Avoid transaction sizes exceed available memory
• Update in smaller chunks e.g. 50k updates per tx
• (2/3G heap per 1M updates in a tx)
• 3 options
◦ Send chunks (list of dicts) from client: UNWIND $data AS row …
◦ CALL {} IN TRANSACTIONS [OF 50000 ROWS]
◦ APOC apoc.periodic.iterate
• USING PERIODIC COMMIT is deprecated
• Avoid the Eager

41
7. Avoid the Eager
• To isolate dependent reads from writes
• Cypher injects an Eager operator that
• pulls ALL rows through that happen before the operator
• Which can use a lot of memory,
• and can disable transaction batching
• Avoided through:
• Multi-pass

42
9. Might help to
• To isolate dependent reads from writes
• Cypher injects an Eager operator that
• pulls ALL rows through that happen before the operator
• Which can use a lot of memory, and can disable batching

43
10. User Defined
Procedures & Functions

44
10. User Defined Procedures & Functions
• Lego blocks to combine inside a Cypher query
• Procedure libraries like APOC cover a wide range of utilities (have a look)
• Refactoring
• Data integration
• Import
• Map/collection/datetime functions
• SHOW PROCEDURES / FUNCTIONS / call apoc.help(„text“)
• CALL procedure(arg1, arg2) YIELD col1, col2

45
10. User Defined Procedures & Functions

46
11. Upcoming – Path
Filters

47
11. Upcoming – Path Filters
• Working towards GQL support
• Inline WHERE predicates for nodes, relationships (5.0)
• Future: predicates for pattern
• Quantifier for patterns / subpatterns

48
Learn more
1 2
Docs
neo4j.com/docs/cypher-
manual/current/query-tuning/
Course
neo4j.com/graphacademy/training-cqt-40/
3 4
Refcard
neo4j.com/docs/cypher-refcard
Community
community.neo4j.com

49
Q&A

50
Thank you!
Contact us at
sales@neo4j.com

51
Split 3rds - bullets
When using bullets on the left ,
select all the bullets and use:
Add space after list item
to visually separate them
Example of bullets:
• Short point
here not more
than 3 lines
• Another point
which is not too
long
• Short and
sweet

52
Neo4j Aura Enterprise is allowing us to realize our
vision to build a knowledge graph that unifies all
product knowledge to improve our retail
recommendations, search and merchandising.”
“
- Name and position, Company

53
Special slides examples
1 2
Arial
Extra Bold 86pt
for big numbers
Arial
Normal 14pt for
small text
3 4
Arial
Extra Bold 86pt
for big numbers
Arial
Normal 14pt for
small text

54
Thing to compare 2
Thing to compare 1
• Comparison 1
• Comparison 2
◦ Details
• Comparison 1
• Comparison 2
◦ Details

55
1. Use Media Slide from Master
2. Image is cropped at width
4.62in, so it takes up half of
the slide (and aligned with
blue rectangle on the left)
Media

56
Image Placeholder

57

58
Your Logo Here

59
59
Your Logo Here

GraphConnect 2022 - Top 10 Cypher Tuning Tips & Tricks.pptx

More Related Content

What's hot (20)

Similar to GraphConnect 2022 - Top 10 Cypher Tuning Tips & Tricks.pptx (20)

More from jexp (20)

Recently uploaded (20)

GraphConnect 2022 - Top 10 Cypher Tuning Tips & Tricks.pptx

Editor's Notes