MongoDB Best Practices

MongoDB Best Practices
Jay Runkel
Principal Solutions Architect
jay.runkel@mongodb.com
@jayrunkel

About Me
• Solution Architect
• Part of Sales Organization
• Work with many organizations new to MongoDB

Everyone Loves MongoDB’s Flexibility
• Document Model
• Dynamic Schema
• Powerful Query Language
• Secondary Indexes

Sometimes Organizations Struggle with
Performance

Good News!
• Poor Performance Usually Due to Common (and often simple) mistakes

Agenda
• Quick MongoDB Introduction
• Best Practices
1. Hardware/OS
2. Schema/Queries
3. Loading Data

Document Data Model
Relational MongoDB
{
first_name: ‘Paul’,
surname: ‘Miller’,
city: ‘London’,
location: [45.123,47.232],
cars: [
{ model: ‘Bentley’,
year: 1973,
value: 100000, … },
{ model: ‘Rolls Royce’,
year: 1965,
value: 330000, … }
]
}

Documents are Rich Data Structures
{
cell: 447557505611,
city: ‘London’,
location: [45.123,47.232],
Profession: [‘banking’, ‘finance’, ‘trader’],
cars: [
year: 1973,
value: 100000, … },
year: 1965,
value: 330000, … }
Fields can contain an array of
sub-documents
Fields
Typed fields
Fields can
contain arrays

Do More With Your Data
{
city: ‘London’,
location: [45.123,47.232],
cars: [
year: 1973,
value: 100000, … },
year: 1965,
value: 330000, … }
}
}
Rich Queries
Find everybody in London
with a car built between
1970 and 1980
Geospatial
Find all of the car owners
within 5km of Trafalgar Sq.
Text Search
Find all the cars described
as having leather seats
Aggregation
Calculate the average value
of Paul’s car collection
Map Reduce
What is the ownership
pattern of colors by
geography over time?
(is purple trending up in
China?)

Automatic Sharding
Three types: hash-based, range-based, location-
aware
Increase or decrease capacity as you go
Automatic balancing

Query Routing
Multiple query optimization
models
Each sharding option
appropriate for different
apps
mongos

Replica Sets
Replica Set – 2 to 50 copies
Self-healing shard
Data Center Aware
Addresses availability considerations:
High Availability
Disaster Recovery
Maintenance
Workload Isolation: operational & analytics

Assumptions
MongoDB 3.0 or 3.2

Storage Engine Architecture in 3.2
Content
Repo
IoT Sensor
Backend
Ad Service
Customer
Analytics
Archive
MongoDB Query Language (MQL) + Native Drivers
MongoDB Document Data Model
WT MMAP
Supported in MongoDB 3.2
Management
Security
In-memory
(beta)
Encrypted 3rd party

Best Practices
Hardware/Operating System

Servers
• Specifications Good Fit For MongoDB?
• Correct Number of Servers?
• Properly Configured?

What Type of Servers
• RAM
– 64  256 GB+
• Fast IO Systems
– RAID-10/SSDs
• Many cores
– Compress/Uncompress
– Encrypt/Decrypt
– Aggregation queries

What about a SAN?
• Mostly Random Disk Access
• IOPS
• Need dedicated IOPS or performance will vary
• Configure your SAN properly
• Suitability of any IO system will depend upon IOPS

How Many Servers Do I Need?
• How Many Shards Do I Need?

MongoDB cluster sizing at 30,000 ft
• Disk Space
• RAM
• Query Throughput

• Sum of disk space across shards > greater than required storage size
Disk Space: How Many Shards Do I Need?

• Sum of disk space across shards > greater than required storage size
Disk Space: How Many Shards Do I Need?
Example
Data Size = 9 TB
WiredTiger Compression Ratio: .33
Storage size = 3 TB
Server disk capacity = 2 TB
2 Shards Required

• Working set should fit in RAM
– Sum of RAM across shards > Working Set
• WorkSet = Indexes plus the set of documents accessed frequently
• WorkSet in RAM 
– Shorter latency
– Higher Throughput
RAM: How Many Shards Do I Need?

• Measuring Index Size
– db.coll.stats() – index size of collection
• Estimate frequently accessed documents
– Ex: total size of documents accessed
per day

• Measuring Index Size
– db.coll.stats() – index size of collection
• Estimate frequently accessed documents
– Ex: total size of documents accessed
per day
Example
Working Set = 428 GB
Server RAM = 128 GB
428/128 = 3.34
4 Shards Required

• Measure max sustained query rate of a single server (with replication)
– build a prototype and measure
• Assume sharding overhead of 20-30%
Query Rate: How Many Shards Do I Need?

• Measure max sustained query rate of a single server (with replication)
– build a prototype and measure
• Assume sharding overhead of 20-30%
Query Rate: How Many Shards Do I Need?
Example
Require: 50K ops/sec
Prototype performance: 20 ops/sec (1
replica set)
4 Shards Required: 80 ops/sec * .7 =
56K ops/sec

Configure Them Properly
• Default OS Settings Often Don’t Provide Optimal Performance
• See MongoDB Production Notes
– https://docs.mongodb.org/manual/administration/production-notes
• Also Review:
– Amazon EC2: https://docs.mongodb.org/ecosystem/platforms/amazon-ec2/
– Azure: https://docs.mongodb.org/ecosystem/platforms/windows-azure/

Server/OS Configuration
• Server configuration recommendations
– XFS
– Turn off atime and diratime
– NOOP scheduler
– File descriptor limits
– Disable transparent huge pages and NUMA
– Read ahead of 32
– Separate data volumes for data files, the journal, and the log.
– Change the default TCP keepalive time to 300 seconds.

These are important
• Ignore them and your performance may suffer
• The first 100 lines of the MongoDB logs identifies
suboptimal OS settings

Don’t Use a Relational Schema

Taylor MongoDB Schema toApplication
Workload
• Design schema to provide good query performance
• Schema design will impact required number of shards!
Application
Query Workload
{
Name: “john”
Height: 12
Address: {…}
}
db.cust.find({…})
db.cust.aggregate({…})

Compare Alternative Schemas
• Build a spreadsheet
• Calculate # of shards for each schema
• Estimate query performance
– # of documents
– # of inserts
– # of deletes
– Required indexes
– Number of documents inspected
– Number of documents sent across network

Modeling Decisions
• Referencing vs. Embedding
• Aggregating data by device, customer, product, etc.

Referencing
Procedure
{
"_id" : 333,
"date" : "2003-02-09T05:00:00"),
"hospital" : “County Hills”,
"patient" : “John Doe”,
"physician" : “Stephen Smith”,
"type" : ”Chest X-ray",
”result" : 134
}
Results
{
“_id” : 134
"type" : "txt",
"size" : NumberInt(12),
"content" : {
value1: 343,
value2: “abc”,
…
}
}

EmbeddingProcedure
{
"_id" : 333,
"date" : "2003-02-09T05:00:00"),
"hospital" : “County Hills”,
"patient" : “John Doe”,
"physician" : “Stephen Smith”,
"type" : ”Chest X-ray",
”result" : {
"type" : "txt",
"size" : NumberInt(12),
"content" : {
value1: 343,
value2: “abc”,
…
}
}
}

Embedding
• Advantages
– Retrieve all relevant information in a single query/document
– Avoid implementing joins in application code
– Update related information as a single atomic operation
• MongoDB doesn’t offer multi-document transactions
• Limitations
– Large documents mean more overhead if most fields are not relevant
– Might mean replicating data
– 16 MB document size limit

Referencing
• Advantages
– Smaller documents
– Less likely to reach 16 MB document limit
– Infrequently accessed information not accessed on every query
– No duplication of data
• Limitations
– Two queries required to retrieve information
– Cannot update related information atomically

{
_id: 2,
first: “Joe”,
last: “Patient”,
addr: { …},
procedures: [
{
id: 12345,
date: 2015-02-15,
type: “Cat scan”,
…},
{
id: 12346,
date: 2015-02-15,
type: “blood test”,
…}]
}
Patients
Embed
One-to-Many & Many-to-Many Relationships
{
_id: 2,
first: “Joe”,
last: “Patient”,
addr: { …},
procedures: [12345, 12346]}
{
_id: 12345,
date: 2015-02-15,
type: “Cat scan”,
…}
{
_id: 12346,
date: 2015-02-15,
type: “blood test”,
…}
Patients
Reference
Procedures

Schema Alternatives – Do the math?
• How complex queries?
• How much hardware/shards will I need?

Vital Sign Monitoring Device
Vital Signs Measured:
• Blood Pressure
• Pulse
• Blood Oxygen Levels
Produces data at regular intervals
• Once per minute

We have a hospital(s) of devices

Data From Vital Signs Monitoring Device
{
deviceId: 123456,
spO2: 88,
pulse: 74,
bp: [128, 80],
ts: ISODate("2013-10-16T22:07:00.000-0500")
}
• One document per minute per device
• Relational approach

Document Per Hour (By minute)
{
deviceId: 123456,
spO2: { 0: 88, 1: 90, …, 59: 92},
pulse: { 0: 74, 1: 76, …, 59: 72},
bp: { 0: [122, 80], 1: [126, 84], …, 59: [124, 78]},
ts: ISODate("2013-10-16T22:00:00.000-0500")
}
• Store per-minute data at the hourly level
• Update-driven workload
• 1 document per device per hour

Characterizing Write Differences
• Example: data generated every minute
• Recording the data for 1 patient for 1 hour:
Document Per Event
60 inserts
Document Per Hour
1 insert, 59 updates

Characterizing Read Differences
• Want to graph 24 hour of vital signs for a patient:
• Read performance is greatly improved
Document Per Event
1440 reads
Document Per Hour
24 reads

Characterizing Memory and Storage
Differences
Document Per Minute Document Per Hour
Number Documents 52.6 B 876 M
Total Index Size 6364 GB 106 GB
_id index 1468 GB 24.5 GB
{ts: 1, deviceId: 1} 4895 GB 81.6 GB
Document Size 92 Bytes 758 Bytes
Database Size 4503 GB 618 GB
• 100K Devices
• 1 years worth of data

Differences
{ts: 1, deviceId: 1} 4895 GB 81.6 GB
• 100K Devices
100000 *
365 * 24 *
60
100000 *
365 * 24

Differences
{ts: 1, deviceId: 1} 4895 GB 81.6 GB
• 100K Devices
100000 *
365 * 24 *
60 * 130
100000 *
365 * 24 *
130

Differences
{ts: 1, deviceId: 1} 4895 GB 81.6 GB
• 100K Devices
100000 *
365 * 24 *
60 * 92
100000 *
365 * 24 *
758

Rule of Thumb
• To saturate a MongoDB cluster 
– loader hardware ~= mongodb hardware
• Many threads
• Many mongos

Loader Architecture
loader
mongos
primary
primary
primary
secondary
secondary
secondary
secondary
secondary
secondary

Loader Architecture
loader
mongos
primary
primary
primary
secondary
secondary
secondary
secondary
secondary
secondary
Where are the bottlenecks?

Loader Architecture
loader (8)
mongos (4)
primary
primary
primary
secondary
secondary
secondary
secondary
secondary
secondary
loader (8)
mongos (4)
loader (8)
mongos (4)
Use many
threads
Use
multiple
loader
servers

When Sharding
• If you care about initial performance, you must pre-split
• Otherwise, initial performance will be slow
• (hash sharding automatically presplits collection)

Without presplitting
Shard 1 Shard 2 Shard 3 Shard 4
-∞ … ∞
• sh.shardCollection(“records.patients”, {zipcode : 1})

-∞ … 11305
• 64K chunks
• Splitting will occur quickly
• Balancing occurs much more slowly
• The entire query workload  Shard 1
11306 … 44506
44507 … ∞

-∞ … 11305
11306 … 44506
44507 … ∞
Loader
mongos

Split collection
• Split and distribute empty chunks before loading any data
• Evenly distribute query load across cluster
-∞ … 08333
08334 … 16667
16668 … 25000
25001… 33334
33335 … 41668
41669 … 50000
50001 … 58334
58335 … 66668
66669 … 75000
75001 … 83334
88335 … 96668
96669 … 99999

Split collection
-∞ … 08333
08334 … 16667
16668 … 25000
25001… 33334
33335 … 41668
41669 … 50000
50001 … 58334
58335 … 66668
66669 … 75000
75001 … 83334
88335 … 96668
96669 … 99999
Loader
mongos

Best Practices
1. Use servers with specifications that will provide good MongoDB performance
– 64+ GB RAM, many cores, many IOPS (RAID-10/SSDs)
2. Calculate How Many Shards?
1. Calculate required RAM and Disk Space
2. Build a prototype to determine the ops/sec capacity of a server
3. Do the math
3. Configure OS for Optimal MongoDB Performance
– See MongoDB Production Notes
– Review logs for warnings (Don’t ignore)

Best Practices (cont.)
4. Create a Document Schema
– Denormalized
5. Tailor schema to application workload
– Use application queries to guide schema design decisions
– Consider alternative schemas
– Compare cluster size (# of shards) and performance
– Build a spreadsheet

Best Practices
6. Loading Data
– Loader Hardware ~= MongoDB hardware
– Many threads
– Many mongos
7. Pre-split
– Ensure query workload is evenly distributed across the cluster from the start

Questions?
jay.runkel@mongodb.com
@jayrunkel

MongoDB Best Practices

In this document