Performance Tuning

This tutorial shows you essential techniques for getting fast reads and writes in CockroachDB, starting with a single-region deployment and expanding into multiple regions.

For a comprehensive list of tuning recommendations, only some of which are demonstrated here, see SQL Performance Best Practices.

Overview

Topology

You'll start with a 3-node CockroachDB cluster in a single Google Compute Engine (GCE) zone, with an extra instance for running a client application workload:

Note:

Within a single GCE zone, network latency between instances should be sub-millisecond.

You'll then scale the cluster to 9 nodes running across 3 GCE regions, with an extra instance in each region for a client application workload:

To reproduce the performance demonstrated in this tutorial:

• For each CockroachDB node, you'll use the n1-standard-4 machine type (4 vCPUs, 15 GB memory) with the Ubuntu 16.04 OS image and a local SSD disk.
• For running the client application workload, you'll use smaller instances, such as n1-standard-1.

Schema

Your schema and data will be based on our open-source, fictional peer-to-peer ride-sharing application,MovR.

A few notes about the schema:

• There are just three self-explanatory tables: In essence, users represents the people registered for the service, vehicles represents the pool of vehicles for the service, and rides represents when and where users have participated.
• Each table has a composite primary key, with city being first in the key. Although not necessary initially in the single-region deployment, once you scale the cluster to multiple regions, these compound primary keys will enable you to geo-partition data at the row level by city. As such, this tutorial demonstrates a schema designed for future scaling.
• The IMPORT feature you'll use to import the data does not support foreign keys, so you'll import the data without foreign key constraints. However, the import will create the secondary indexes required to add the foreign keys later.
• The rides table contains both city and the seemingly redundant vehicle_city. This redundancy is necessary because, while it is not possible to apply more than one foreign key constraint to a single column, you will need to apply two foreign key constraints to the rides table, and each will require city as part of the constraint. The duplicate vehicle_city, which is kept in synch with city via a CHECK constraint, lets you overcome this limitation.

Important concepts

To understand the techniques in this tutorial, and to be able to apply them in your own scenarios, it's important to first understand how reads and writes work in CockroachDB. Review that document before getting started here.

Single-region deployment

Step 1. Configure your network

CockroachDB requires TCP communication on two ports:

• 26257 (tcp:26257) for inter-node communication (i.e., working as a cluster)
• 8080 (tcp:8080) for accessing the Admin UI
  Since GCE instances communicate on their internal IP addresses by default, you do not need to take any action to enable inter-node communication. However, to access the Admin UI from your local network, you must create a firewall rule for your project:

Note:

The tag feature will let you easily apply the rule to your instances.

Step 2. Create instances

You'll start with a 3-node CockroachDB cluster in the us-east1-b GCE zone, with an extra instance for running a client application workload.

• Create 3 instances for your CockroachDB nodes. While creating each instance:

  • Select the us-east1-b zone.
  • Use the n1-standard-4 machine type (4 vCPUs, 15 GB memory).
  • Use the Ubuntu 16.04 OS image.
  • Create and mount a local SSD.
  • To apply the Web UI firewall rule you created earlier, click Management, disk, networking, SSH keys, select the Networking tab, and then enter cockroachdb in the Network tags field.

• Note the internal and external IP addresses of each n1-standard-4 instance. You'll need these addresses when generating security certificates and when starting the CockroachDB nodes.

• Create a separate instance for running a client application workload, also in the us-east1-b zone. This instance can be smaller, such as n1-standard-1.

Step 3. Start a 3-node cluster

Generate security certificates

You can use either cockroach cert commands or openssl commands to generate security certificates. This section features the cockroach cert commands.

Locally, you'll need to create the following certificates and keys:

• A certificate authority (CA) key pair (ca.crt and ca.key).
• A node key pair for each node, issued to its IP addresses and any common names the machine uses.
• A client key pair for the root user. You'll use this when running you client application workload as well as some cockroach client commands.

Tip:

As mentioned above, before beginning, it's useful to collect each instance's internal and external IP addresses, as well as any server names you want to issue certificates for.

• Install CockroachDB on your local machine, if you haven't already.

• Create two directories:

$ mkdir certs

$ mkdir my-safe-directory

• certs: You'll generate your CA certificate and all node and client certificates and keys in this directory and then upload some of the files to your nodes.
• my-safe-directory: You'll generate your CA key in this directory and then reference the key when generating node and client certificates. After that, you'll keep the key safe and secret; you will not upload it to your nodes.
  • Create the CA certificate and key:

$ cockroach cert create-ca \
--certs-dir=certs \
--ca-key=my-safe-directory/ca.key

• Create the certificate and key for the first node, issued to all common names you might use to refer to the node:

$ cockroach cert create-node \
<node1 internal IP address> \
<node1 external IP address> \
<node1 hostname>  \
<other common names for node1> \
localhost \
127.0.0.1 \
--certs-dir=certs \
--ca-key=my-safe-directory/ca.key

• Upload certificates to the first instance:

# Create the certs directory:
$ ssh <username>@<node1 address> "mkdir certs"

# Upload the CA certificate and node certificate and key:
$ scp certs/ca.crt \
certs/node.crt \
certs/node.key \
<username>@<node1 address>:~/certs

• Delete the local copy of the node certificate and key:

$ rm certs/node.crt certs/node.key

Note:

This is necessary because the certificates and keys for additional nodes will also be named node.crt and node.key. As an alternative to deleting these files, you can run the next cockroach cert create-node commands with the —overwrite flag.

• Create the certificate and key for the second node, issued to all common names you might use to refer to the node:

$ cockroach cert create-node \
<node2 internal IP address> \
<node2 external IP address> \
<node2 hostname>  \
<other common names for node2> \
localhost \
127.0.0.1 \
--certs-dir=certs \
--ca-key=my-safe-directory/ca.key

• Upload certificates to the second instance:

# Create the certs directory:
$ ssh <username>@<node2 address> "mkdir certs"

# Upload the CA certificate and node certificate and key:
$ scp certs/ca.crt \
certs/node.crt \
certs/node.key \
<username>@<node2 address>:~/certs

• Repeat steps 6 - 8 for the third node.

• Create a client certificate and key for the root user:

$ cockroach cert create-client \
root \
--certs-dir=certs \
--ca-key=my-safe-directory/ca.key

• Upload certificates to the fourth instance, the one from which you will run a sample workload:

# Create the certs directory:
$ ssh <username>@<instance4 address> "mkdir certs"

# Upload the CA certificate and client certificate and key:
$ scp certs/ca.crt \
certs/client.root.crt \
certs/client.root.key \
<username>@<instance4 address>:~/certs

Note:

On accessing the Admin UI in a later step, your browser will consider the CockroachDB-created certificate invalid and you’ll need to click through a warning message to get to the UI. You can avoid this issue by using a certificate issued by a public CA.

Start the nodes

• SSH to the first n1-standard-4 instance.

• Download the CockroachDB archive for Linux, extract the binary, and copy it into the PATH:

$ wget -qO- https://binaries.cockroachdb.com/cockroach-v19.1.0.linux-amd64.tgz \
| tar  xvz

$ sudo cp -i cockroach-v19.1.0.linux-amd64/cockroach /usr/local/bin

• Run the cockroach start command:

$ cockroach start \
--certs-dir=certs \
--advertise-host=<node1 internal address> \
--join=<node1 internal address>:26257,<node2 internal address>:26257,<node3 internal address>:26257 \
--locality=cloud=gce,region=us-east1,zone=us-east1-b \
--cache=.25 \
--max-sql-memory=.25 \
--background

• Repeat steps 1 - 3 for the other two n1-standard-4 instances. Be sure to adjust the —advertise-addr flag each time.

Initialize the cluster

• SSH to the fourth instance, the one not running a CockroachDB node.

• Download the CockroachDB archive for Linux, and extract the binary:

$ wget -qO- https://binaries.cockroachdb.com/cockroach-v19.1.0.linux-amd64.tgz \
| tar  xvz

• Copy the binary into the PATH:

$ sudo cp -i cockroach-v19.1.0.linux-amd64/cockroach /usr/local/bin

• Run the cockroach init command:

$ cockroach init --certs-dir=certs --host=<address of any node>

Each node then prints helpful details to the standard output, such as the CockroachDB version, the URL for the Web UI, and the SQL URL for clients.

Step 4. Import the Movr dataset

Now you'll import Movr data representing users, vehicles, and rides in 3 eastern US cities (New York, Boston, and Washington DC) and 3 western US cities (Los Angeles, San Francisco, and Seattle).

• Still on the fourth instance, start the built-in SQL shell, pointing it at one of the CockroachDB nodes:

$ cockroach sql --certs-dir=certs --host=<address of any node>

• Create the movr database and set it as the default:

> CREATE DATABASE movr;

> SET DATABASE = movr;

• Use the IMPORT statement to create and populate the users, vehicles, and rides tables:

> IMPORT TABLE users (
    id UUID NOT NULL,
    city STRING NOT NULL,
    name STRING NULL,
    address STRING NULL,
    credit_card STRING NULL,
    CONSTRAINT "primary" PRIMARY KEY (city ASC, id ASC)
)
CSV DATA (
    'https://s3-us-west-1.amazonaws.com/cockroachdb-movr/datasets/perf-tuning/users/n1.0.csv'
);

        job_id       |  status   | fraction_completed | rows | index_entries | system_records | bytes
+--------------------+-----------+--------------------+------+---------------+----------------+--------+
  390345990764396545 | succeeded |                  1 | 1998 |             0 |              0 | 241052
(1 row)
Time: 2.882582355s

> IMPORT TABLE vehicles (
    id UUID NOT NULL,
    city STRING NOT NULL,
    type STRING NULL,
    owner_id UUID NULL,
    creation_time TIMESTAMP NULL,
    status STRING NULL,
    ext JSON NULL,
    mycol STRING NULL,
    CONSTRAINT "primary" PRIMARY KEY (city ASC, id ASC),
    INDEX vehicles_auto_index_fk_city_ref_users (city ASC, owner_id ASC)
)
CSV DATA (
    'https://s3-us-west-1.amazonaws.com/cockroachdb-movr/datasets/perf-tuning/vehicles/n1.0.csv'
);

        job_id       |  status   | fraction_completed | rows  | index_entries | system_records |  bytes
+--------------------+-----------+--------------------+-------+---------------+----------------+---------+
  390346109887250433 | succeeded |                  1 | 19998 |         19998 |              0 | 3558767
(1 row)
Time: 5.803841493s

> IMPORT TABLE rides (
    id UUID NOT NULL,
    city STRING NOT NULL,
    vehicle_city STRING NULL,
    rider_id UUID NULL,
    vehicle_id UUID NULL,
    start_address STRING NULL,
    end_address STRING NULL,
    start_time TIMESTAMP NULL,
    end_time TIMESTAMP NULL,
    revenue DECIMAL(10,2) NULL,
    CONSTRAINT "primary" PRIMARY KEY (city ASC, id ASC),
    INDEX rides_auto_index_fk_city_ref_users (city ASC, rider_id ASC),
    INDEX rides_auto_index_fk_vehicle_city_ref_vehicles (vehicle_city ASC, vehicle_id ASC),
    CONSTRAINT check_vehicle_city_city CHECK (vehicle_city = city)
)
CSV DATA (
    'https://s3-us-west-1.amazonaws.com/cockroachdb-movr/datasets/perf-tuning/rides/n1.0.csv',
    'https://s3-us-west-1.amazonaws.com/cockroachdb-movr/datasets/perf-tuning/rides/n1.1.csv',
    'https://s3-us-west-1.amazonaws.com/cockroachdb-movr/datasets/perf-tuning/rides/n1.2.csv',
    'https://s3-us-west-1.amazonaws.com/cockroachdb-movr/datasets/perf-tuning/rides/n1.3.csv',
    'https://s3-us-west-1.amazonaws.com/cockroachdb-movr/datasets/perf-tuning/rides/n1.4.csv',
    'https://s3-us-west-1.amazonaws.com/cockroachdb-movr/datasets/perf-tuning/rides/n1.5.csv',
    'https://s3-us-west-1.amazonaws.com/cockroachdb-movr/datasets/perf-tuning/rides/n1.6.csv',
    'https://s3-us-west-1.amazonaws.com/cockroachdb-movr/datasets/perf-tuning/rides/n1.7.csv',
    'https://s3-us-west-1.amazonaws.com/cockroachdb-movr/datasets/perf-tuning/rides/n1.8.csv',
    'https://s3-us-west-1.amazonaws.com/cockroachdb-movr/datasets/perf-tuning/rides/n1.9.csv'
);

        job_id       |  status   | fraction_completed |  rows  | index_entries | system_records |   bytes
+--------------------+-----------+--------------------+--------+---------------+----------------+-----------+
  390346325693792257 | succeeded |                  1 | 999996 |       1999992 |              0 | 339741841
(1 row)
Time: 44.620371424s

Tip:

You can observe the progress of imports as well as all schema change operations (e.g., adding secondary indexes) on the Jobs page of the Web UI.

• Logically, there should be a number of foreign key relationships between the tables:

Referencing columnsReferenced columnsvehicles.city, vehicles.owner_idusers.city, users.idrides.city, rides.rider_idusers.city, users.idrides.vehicle_city, rides.vehicle_idvehicles.city, vehicles.id

As mentioned earlier, it wasn't possible to put these relationships in place during IMPORT, but it was possible to create the required secondary indexes. Now, let's add the foreign key constraints:

> ALTER TABLE vehicles
ADD CONSTRAINT fk_city_ref_users
FOREIGN KEY (city, owner_id)
REFERENCES users (city, id);

> ALTER TABLE rides
ADD CONSTRAINT fk_city_ref_users
FOREIGN KEY (city, rider_id)
REFERENCES users (city, id);

> ALTER TABLE rides
ADD CONSTRAINT fk_vehicle_city_ref_vehicles
FOREIGN KEY (vehicle_city, vehicle_id)
REFERENCES vehicles (city, id);

• Exit the built-in SQL shell:

> \q

Step 5. Install the Python client

When measuring SQL performance, it's best to run a given statement multiple times and look at the average and/or cumulative latency. For that purpose, you'll install and use a Python testing client.

• Still on the fourth instance, make sure all of the system software is up-to-date:

$ sudo apt-get update && sudo apt-get -y upgrade

• Install the psycopg2 driver:

$ sudo apt-get install python-psycopg2

• Download the Python client:

$ wget https://raw.githubusercontent.com/cockroachdb/docs/master/_includes/v19.1/performance/tuning-secure.py \
&& chmod +x tuning-secure.py

As you'll see below, this client lets you pass command-line flags:

FlagDescription—hostThe IP address of the target node. This is used in the client's connection string.—statementThe SQL statement to execute.—repeatThe number of times to repeat the statement. This defaults to 20.

When run, the client prints the median time in seconds across all repetitions of the statement. Optionally, you can pass two other flags, —time to print the execution time in seconds for each repetition of the statement, and —cumulative to print the cumulative time in seconds for all repetitions. —cumulative is particularly useful when testing writes.

Tip:

To get similar help directly in your shell, use ./tuning-secure.py —help.

Step 6. Test/tune read performance

• Filtering by the primary key
• Filtering by a non-indexed column (full table scan)
• Filtering by a secondary index
• Filtering by a secondary index storing additional columns
• Joining data from different tables
• Using IN (list) with a subquery
• Using IN (list) with explicit values

Filtering by the primary key

Retrieving a single row based on the primary key will usually return in 2ms or less:

$ ./tuning-secure.py \
--host=<address of any node> \
--statement="SELECT * FROM rides WHERE city = 'boston' AND id = '000007ef-fa0f-4a6e-a089-ce74aa8d2276'" \
--repeat=50 \
--times

Result:
['id', 'city', 'vehicle_city', 'rider_id', 'vehicle_id', 'start_address', 'end_address', 'start_time', 'end_time', 'revenue']
['000007ef-fa0f-4a6e-a089-ce74aa8d2276', 'boston', 'boston', 'd66c386d-4b7b-48a7-93e6-f92b5e7916ab', '6628bbbc-00be-4891-bc00-c49f2f16a30b', '4081 Conner Courts\nSouth Taylor, VA 86921', '2808 Willis Wells Apt. 931\nMccoyberg, OH 10303-4879', '2018-07-20 01:46:46.003070', '2018-07-20 02:27:46.003070', '44.25']
Times (milliseconds):
[2.1638870239257812, 1.2159347534179688, 1.0809898376464844, 1.0669231414794922, 1.2650489807128906, 1.1401176452636719, 1.1310577392578125, 1.0380744934082031, 1.199960708618164, 1.0530948638916016, 1.1000633239746094, 1.3430118560791016, 1.104116439819336, 1.0750293731689453, 1.0609626770019531, 1.088857650756836, 1.1639595031738281, 1.2559890747070312, 1.1899471282958984, 1.0449886322021484, 1.1057853698730469, 1.127004623413086, 0.9729862213134766, 1.1131763458251953, 1.0879039764404297, 1.119852066040039, 1.065969467163086, 1.0371208190917969, 1.1181831359863281, 1.0409355163574219, 1.0859966278076172, 1.1398792266845703, 1.032114028930664, 1.1000633239746094, 1.1360645294189453, 1.146078109741211, 1.329183578491211, 1.1131763458251953, 1.1548995971679688, 0.9977817535400391, 1.1138916015625, 1.085042953491211, 1.0950565338134766, 1.0869503021240234, 1.0170936584472656, 1.0571479797363281, 1.0640621185302734, 1.1110305786132812, 1.1279582977294922, 1.1119842529296875]
Median time (milliseconds):
1.10495090485

Retrieving a subset of columns will usually be even faster:

$ ./tuning-secure.py \
--host=<address of any node> \
--statement="SELECT rider_id, vehicle_id \
FROM rides \
WHERE city = 'boston' AND id = '000007ef-fa0f-4a6e-a089-ce74aa8d2276'" \
--repeat=50 \
--times

Result:
['rider_id', 'vehicle_id']
['d66c386d-4b7b-48a7-93e6-f92b5e7916ab', '6628bbbc-00be-4891-bc00-c49f2f16a30b']
Times (milliseconds):
[2.218961715698242, 1.2569427490234375, 1.3570785522460938, 1.1570453643798828, 1.3251304626464844, 1.3320446014404297, 1.0790824890136719, 1.0139942169189453, 1.0251998901367188, 1.1150836944580078, 1.1949539184570312, 1.2140274047851562, 1.2080669403076172, 1.238107681274414, 1.071929931640625, 1.104116439819336, 1.0230541229248047, 1.0571479797363281, 1.0519027709960938, 1.0688304901123047, 1.0118484497070312, 1.0051727294921875, 1.1889934539794922, 1.0571479797363281, 1.177072525024414, 1.0449886322021484, 1.0669231414794922, 1.004934310913086, 0.9818077087402344, 0.9369850158691406, 1.004934310913086, 1.0461807250976562, 1.0628700256347656, 1.1332035064697266, 1.1780261993408203, 1.0361671447753906, 1.1410713195800781, 1.1188983917236328, 1.026153564453125, 0.9629726409912109, 1.0199546813964844, 1.0409355163574219, 1.0440349578857422, 1.1110305786132812, 1.1761188507080078, 1.508951187133789, 1.2068748474121094, 1.3430118560791016, 1.4159679412841797, 1.3141632080078125]
Median time (milliseconds):
1.09159946442

Filtering by a non-indexed column (full table scan)

You'll get generally poor performance when retrieving a single row based on a column that is not in the primary key or any secondary index:

$ ./tuning-secure.py \
--host=<address of any node> \
--statement="SELECT * FROM users WHERE name = 'Natalie Cunningham'" \
--repeat=50 \
--times

Result:
['id', 'city', 'name', 'address', 'credit_card']
['02cc9e5b-1e91-4cdb-87c4-726b4ea7219a', 'boston', 'Natalie Cunningham', '97477 Lee Path\nKimberlyport, CA 65960', '4532613656695680']
Times (milliseconds):
[33.271074295043945, 4.4689178466796875, 4.18400764465332, 4.327058792114258, 5.700111389160156, 4.509925842285156, 4.525899887084961, 4.294157028198242, 4.516124725341797, 5.700111389160156, 5.105018615722656, 4.5070648193359375, 4.798173904418945, 5.930900573730469, 4.445075988769531, 4.1790008544921875, 4.065036773681641, 4.296064376831055, 5.722999572753906, 4.827976226806641, 4.640102386474609, 4.374980926513672, 4.269123077392578, 4.422903060913086, 4.110813140869141, 4.091024398803711, 4.189014434814453, 4.345178604125977, 5.600929260253906, 4.827976226806641, 4.416942596435547, 4.424095153808594, 4.736185073852539, 4.462003707885742, 4.307031631469727, 5.10096549987793, 4.56690788269043, 4.641056060791016, 4.701137542724609, 4.538059234619141, 4.474163055419922, 4.561901092529297, 4.431009292602539, 4.756927490234375, 4.54401969909668, 4.415035247802734, 4.396915435791016, 5.9719085693359375, 4.543066024780273, 5.830049514770508]
Median time (milliseconds):
4.51302528381

To understand why this query performs poorly, use the SQL client built into the cockroach binary to EXPLAIN the query plan:

$ cockroach sql \
--certs-dir=certs \
--host=<address of any node> \
--database=movr \
--execute="EXPLAIN SELECT * FROM users WHERE name = 'Natalie Cunningham';"

  tree | field |  description
+------+-------+---------------+
  scan |       |
       | table | users@primary
       | spans | ALL
(3 rows)

The row with spans | ALL shows you that, without a secondary index on the name column, CockroachDB scans every row of the users table, ordered by the primary key (city/id), until it finds the row with the correct name value.

Filtering by a secondary index

To speed up this query, add a secondary index on name:

$ cockroach sql \
--certs-dir=certs \
--host=<address of any node> \
--database=movr \
--execute="CREATE INDEX on users (name);"

The query will now return much faster:

$ ./tuning-secure.py \
--host=<address of any node> \
--statement="SELECT * FROM users WHERE name = 'Natalie Cunningham'" \
--repeat=50 \
--times

Result:
['id', 'city', 'name', 'address', 'credit_card']
['02cc9e5b-1e91-4cdb-87c4-726b4ea7219a', 'boston', 'Natalie Cunningham', '97477 Lee Path\nKimberlyport, CA 65960', '4532613656695680']
Times (milliseconds):
[3.545045852661133, 1.4619827270507812, 1.399993896484375, 2.0101070404052734, 1.672983169555664, 1.4941692352294922, 1.4650821685791016, 1.4579296112060547, 1.567840576171875, 1.5709400177001953, 1.4760494232177734, 1.6181468963623047, 1.6210079193115234, 1.6970634460449219, 1.6469955444335938, 1.7261505126953125, 1.7559528350830078, 1.875162124633789, 1.7170906066894531, 1.870870590209961, 1.641988754272461, 1.7061233520507812, 1.628875732421875, 1.6558170318603516, 1.7809867858886719, 1.6698837280273438, 1.8429756164550781, 1.6090869903564453, 1.7080307006835938, 1.74713134765625, 1.6620159149169922, 1.9519329071044922, 1.6849040985107422, 1.7440319061279297, 1.8851757049560547, 1.8699169158935547, 1.7409324645996094, 1.9140243530273438, 1.7828941345214844, 1.7158985137939453, 1.6720294952392578, 1.7750263214111328, 1.7368793487548828, 1.9288063049316406, 1.8749237060546875, 1.7838478088378906, 1.8091201782226562, 1.8210411071777344, 1.7669200897216797, 1.8210411071777344]
Median time (milliseconds):
1.72162055969

To understand why performance improved from 4.51ms (without index) to 1.72ms (with index), use EXPLAIN to see the new query plan:

$ cockroach sql \
--certs-dir=certs \
--host=<address of any node> \
--database=movr \
--execute="EXPLAIN SELECT * FROM users WHERE name = 'Natalie Cunningham';"

     tree    | field |                      description
+------------+-------+-------------------------------------------------------+
  index-join |       |
   ├── scan  |       |
   │         | table | users@users_name_idx
   │         | spans | /"Natalie Cunningham"-/"Natalie Cunningham"/PrefixEnd
   └── scan  |       |
             | table | users@primary
(6 rows)

This shows you that CockroachDB starts with the secondary index (table | users@users_name_idx). Because it is sorted by name, the query can jump directly to the relevant value (spans | /"Natalie Cunningham"-/"Natalie Cunningham"/PrefixEnd). However, the query needs to return values not in the secondary index, so CockroachDB grabs the primary key (city/id) stored with the name value (the primary key is always stored with entries in a secondary index), jumps to that value in the primary index, and then returns the full row.

Thinking back to the earlier discussion of ranges and leaseholders, because the users table is small (under 64 MiB), the primary index and all secondary indexes are contained in a single range with a single leaseholder. If the table were bigger, however, the primary index and secondary index could reside in separate ranges, each with its own leaseholder. In this case, if the leaseholders were on different nodes, the query would require more network hops, further increasing latency.

Filtering by a secondary index storing additional columns

When you have a query that filters by a specific column but retrieves a subset of the table's total columns, you can improve performance by storing those additional columns in the secondary index to prevent the query from needing to scan the primary index as well.

For example, let's say you frequently retrieve a user's name and credit card number:

$ ./tuning-secure.py \
--host=<address of any node> \
--statement="SELECT name, credit_card FROM users WHERE name = 'Natalie Cunningham'" \
--repeat=50 \
--times

Result:
['name', 'credit_card']
['Natalie Cunningham', '4532613656695680']
Times (milliseconds):
[2.8769969940185547, 1.7559528350830078, 1.8100738525390625, 1.8839836120605469, 1.5971660614013672, 1.5900135040283203, 1.7750263214111328, 2.2847652435302734, 1.641988754272461, 1.4967918395996094, 1.4641284942626953, 1.6689300537109375, 1.9679069519042969, 1.8970966339111328, 1.8780231475830078, 1.7609596252441406, 1.68609619140625, 1.9791126251220703, 1.661062240600586, 1.9869804382324219, 1.5938282012939453, 1.8041133880615234, 1.5909671783447266, 1.5878677368164062, 1.7380714416503906, 1.638174057006836, 1.6970634460449219, 1.9309520721435547, 1.992940902709961, 1.8689632415771484, 1.7511844635009766, 2.007007598876953, 1.9829273223876953, 1.8939971923828125, 1.7490386962890625, 1.6179084777832031, 1.6510486602783203, 1.6078948974609375, 1.6129016876220703, 1.67083740234375, 1.786947250366211, 1.7840862274169922, 1.956939697265625, 1.8689632415771484, 1.9350051879882812, 1.789093017578125, 1.9249916076660156, 1.8649101257324219, 1.9619464874267578, 1.7361640930175781]
Median time (milliseconds):
1.77955627441

With the current secondary index on name, CockroachDB still needs to scan the primary index to get the credit card number:

$ cockroach sql \
--certs-dir=certs \
--host=<address of any node> \
--database=movr \
--execute="EXPLAIN SELECT name, credit_card FROM users WHERE name = 'Natalie Cunningham';"

     tree    | field |                      description
+------------+-------+-------------------------------------------------------+
  index-join |       |
   ├── scan  |       |
   │         | table | users@users_name_idx
   │         | spans | /"Natalie Cunningham"-/"Natalie Cunningham"/PrefixEnd
   └── scan  |       |
             | table | users@primary
(6 rows)

Let's drop and recreate the index on name, this time storing the credit_card value in the index:

$ cockroach sql \
--certs-dir=certs \
--host=<address of any node> \
--database=movr \
--execute="DROP INDEX users_name_idx;"

$ cockroach sql \
--certs-dir=certs \
--host=<address of any node> \
--database=movr \
--execute="CREATE INDEX ON users (name) STORING (credit_card);"

Now that credit_card values are stored in the index on name, CockroachDB only needs to scan that index:

$ cockroach sql \
--certs-dir=certs \
--host=<address of any node> \
--database=movr \
--execute="EXPLAIN SELECT name, credit_card FROM users WHERE name = 'Natalie Cunningham';"

  tree | field |                      description
+------+-------+-------------------------------------------------------+
  scan |       |
       | table | users@users_name_idx
       | spans | /"Natalie Cunningham"-/"Natalie Cunningham"/PrefixEnd
(3 rows)

This results in even faster performance, reducing latency from 1.77ms (index without storing) to 0.99ms (index with storing):

$ ./tuning-secure.py \
--host=<address of any node> \
--statement="SELECT name, credit_card FROM users WHERE name = 'Natalie Cunningham'" \
--repeat=50 \
--times

Result:
['name', 'credit_card']
['Natalie Cunningham', '4532613656695680']
Times (milliseconds):
[1.8029212951660156, 0.9858608245849609, 0.9548664093017578, 0.8459091186523438, 0.9710788726806641, 1.1639595031738281, 0.8571147918701172, 0.8800029754638672, 0.8509159088134766, 0.8771419525146484, 1.1739730834960938, 0.9100437164306641, 1.1181831359863281, 0.9679794311523438, 1.0800361633300781, 1.02996826171875, 1.2090206146240234, 1.0440349578857422, 1.210927963256836, 1.0418891906738281, 1.1951923370361328, 0.9548664093017578, 1.0848045349121094, 0.9748935699462891, 1.15203857421875, 1.0280609130859375, 1.0819435119628906, 0.9641647338867188, 1.0979175567626953, 0.9720325469970703, 1.0638236999511719, 0.9410381317138672, 1.0039806365966797, 1.207113265991211, 0.9911060333251953, 1.0039806365966797, 0.9810924530029297, 0.9360313415527344, 0.9589195251464844, 1.0609626770019531, 0.9949207305908203, 1.0139942169189453, 0.9899139404296875, 0.9818077087402344, 0.9679794311523438, 0.8809566497802734, 0.9558200836181641, 0.8878707885742188, 1.0380744934082031, 0.8897781372070312]
Median time (milliseconds):
0.990509986877

Joining data from different tables

Secondary indexes are crucial when joining data from different tables as well.

For example, let's say you want to count the number of users who started rides on a given day. To do this, you need to use a join to get the relevant rides from the rides table and then map the rider_id for each of those rides to the corresponding id in the users table, counting each mapping only once:

$ ./tuning-secure.py \
--host=<address of any node> \
--statement="SELECT count(DISTINCT users.id) \
FROM users \
INNER JOIN rides ON rides.rider_id = users.id \
WHERE start_time BETWEEN '2018-07-20 00:00:00' AND '2018-07-21 00:00:00'" \
--repeat=50 \
--times

Result:
['count']
['1998']
Times (milliseconds):
[1443.5458183288574, 1546.0000038146973, 1563.858985900879, 1530.3218364715576, 1574.7389793395996, 1572.7760791778564, 1566.4539337158203, 1595.655918121338, 1588.2930755615234, 1567.6488876342773, 1564.5530223846436, 1573.4570026397705, 1581.406831741333, 1587.864875793457, 1575.7901668548584, 1565.0341510772705, 1519.8209285736084, 1599.7698307037354, 1612.4188899993896, 1582.5250148773193, 1604.076862335205, 1596.8739986419678, 1569.6821212768555, 1583.7080478668213, 1549.9720573425293, 1563.5790824890137, 1555.6750297546387, 1577.6000022888184, 1582.3569297790527, 1568.8848495483398, 1580.854892730713, 1566.9701099395752, 1578.8500308990479, 1592.677116394043, 1549.3559837341309, 1561.805009841919, 1561.812162399292, 1543.4870719909668, 1523.3290195465088, 1583.9049816131592, 1565.9120082855225, 1575.1979351043701, 1581.1400413513184, 1616.6048049926758, 1602.9179096221924, 1583.8429927825928, 1570.2300071716309, 1573.2421875, 1558.588981628418, 1548.7489700317383]
Median time (milliseconds):
1573.00913334

To understand what's happening, use EXPLAIN to see the query plan:

$ cockroach sql \
--certs-dir=certs \
--host=<address of any node> \
--database=movr \
--execute="EXPLAIN SELECT count(DISTINCT users.id) \
FROM users \
INNER JOIN rides ON rides.rider_id = users.id \
WHERE start_time BETWEEN '2018-07-20 00:00:00' AND '2018-07-21 00:00:00';"

         tree         |    field    |     description
+---------------------+-------------+----------------------+
  group               |             |
   │                  | aggregate 0 | count(DISTINCT id)
   │                  | scalar      |
   └── render         |             |
        └── join      |             |
             │        | type        | inner
             │        | equality    | (id) = (rider_id)
             ├── scan |             |
             │        | table       | users@users_name_idx
             │        | spans       | ALL
             └── scan |             |
                      | table       | rides@primary
                      | spans       | ALL
(13 rows)

Reading from bottom up, you can see that CockroachDB does a full table scan (spans | ALL) first on rides to get all rows with a start_time in the specified range and then does another full table scan on users to find matching rows and calculate the count.

Given that the rides table is large, its data is split across several ranges. Each range is replicated and has a leaseholder. At least some of these leaseholders are likely located on different nodes. This means that the full table scan of rides involves several network hops to various leaseholders before finally going to the leaseholder for users to do a full table scan there.

To track this specifically, let's use the SHOW EXPERIMENTAL_RANGES statement to find out where the relevant leaseholders reside for rides and users:

$ cockroach sql \
--certs-dir=certs \
--host=<address of any node> \
--database=movr \
--execute="SHOW EXPERIMENTAL_RANGES FROM TABLE rides;"

                                  start_key                                  |                                  end_key                                   | range_id | replicas | lease_holder
+----------------------------------------------------------------------------+----------------------------------------------------------------------------+----------+----------+--------------+
  NULL                                                                       | /"boston"/"\xfe\xdd?\xbb4\xabOV\x84\x00M\x89#-a6"/PrefixEnd                |       34 | {1,2,3}  |            2
  /"boston"/"\xfe\xdd?\xbb4\xabOV\x84\x00M\x89#-a6"/PrefixEnd                | /"los angeles"/"<\x12\xe4\xce&\xfdH\u070f?)\xc7\xf92\a\x03"                |       35 | {1,2,3}  |            2
  /"los angeles"/"<\x12\xe4\xce&\xfdH\u070f?)\xc7\xf92\a\x03"                | /"new york"/"0\xa6p\x96\tmOԗ#\xaa\xb7\x90\x12\xe67"/PrefixEnd              |       39 | {1,2,3}  |            2
  /"new york"/"0\xa6p\x96\tmOԗ#\xaa\xb7\x90\x12\xe67"/PrefixEnd              | /"san francisco"/"(m*OM\x15J\xbc\xb6n\xaass\x10\xc4\xff"/PrefixEnd         |       37 | {1,2,3}  |            1
  /"san francisco"/"(m*OM\x15J\xbc\xb6n\xaass\x10\xc4\xff"/PrefixEnd         | /"seattle"/"\x17\xd24\a\xb5\xbdN\x9d\xa1\xd2Dθ^\xe1M"/PrefixEnd            |       40 | {1,2,3}  |            2
  /"seattle"/"\x17\xd24\a\xb5\xbdN\x9d\xa1\xd2Dθ^\xe1M"/PrefixEnd            | /"washington dc"/"\x135\xe5e\x15\xefNۊ\x10)\xba\x19\x04\xff\xdc"/PrefixEnd |       44 | {1,2,3}  |            2
  /"washington dc"/"\x135\xe5e\x15\xefNۊ\x10)\xba\x19\x04\xff\xdc"/PrefixEnd | NULL                                                                       |       46 | {1,2,3}  |            2
(7 rows)

$ cockroach sql \
--certs-dir=certs \
--host=<address of any node> \
--database=movr \
--execute="SHOW EXPERIMENTAL_RANGES FROM TABLE users;"

  start_key | end_key | range_id | replicas | lease_holder
+-----------+---------+----------+----------+--------------+
  NULL      | NULL    |       49 | {1,2,3}  |            2
(1 row)

The results above tell us:

• The rides table is split across 7 ranges, with six leaseholders on node 2 and one leaseholder on node 1.
• The users table is just a single range with its leaseholder on node 2.
  Now, given the WHERE condition of the join, the full table scan of rides, across all of its 7 ranges, is particularly wasteful. To speed up the query, you can create a secondary index on the WHERE condition (rides.start_time) storing the join key (rides.rider_id):

$ cockroach sql \
--certs-dir=certs \
--host=<address of any node> \
--database=movr \
--execute="CREATE INDEX ON rides (start_time) STORING (rider_id);"

Note:

The rides table contains 1 million rows, so adding this index will take a few minutes.

Adding the secondary index reduced the query time from 1573ms to 61.56ms:

$ ./tuning-secure.py \
--host=<address of any node> \
--statement="SELECT count(DISTINCT users.id) \
FROM users \
INNER JOIN rides ON rides.rider_id = users.id \
WHERE start_time BETWEEN '2018-07-20 00:00:00' AND '2018-07-21 00:00:00'" \
--repeat=50 \
--times

Result:
['count']
['1998']
Times (milliseconds):
[66.78199768066406, 63.83800506591797, 65.57297706604004, 63.04502487182617, 61.54489517211914, 61.51890754699707, 60.935020446777344, 61.8891716003418, 60.71019172668457, 64.44311141967773, 64.82601165771484, 61.5849494934082, 62.136173248291016, 62.78491020202637, 62.70194053649902, 61.837196350097656, 64.13102149963379, 62.66903877258301, 71.14315032958984, 61.08808517456055, 58.36200714111328, 60.003042221069336, 58.743953704833984, 59.05413627624512, 60.63103675842285, 60.12582778930664, 61.02705001831055, 62.548160552978516, 61.45000457763672, 65.27113914489746, 60.18996238708496, 59.36002731323242, 60.13298034667969, 59.8299503326416, 59.168100357055664, 65.20915031433105, 60.43219566345215, 58.91895294189453, 58.67791175842285, 59.50117111206055, 59.977054595947266, 65.39011001586914, 62.3931884765625, 69.40793991088867, 61.64288520812988, 66.52498245239258, 69.78988647460938, 60.96601486206055, 57.71303176879883, 61.81192398071289]
Median time (milliseconds):
61.5649223328

To understand why performance improved, again use EXPLAIN to see the new query plan:

$ cockroach sql \
--certs-dir=certs \
--host=<address of any node> \
--database=movr \
--execute="EXPLAIN SELECT count(DISTINCT users.id) \
FROM users \
INNER JOIN rides ON rides.rider_id = users.id \
WHERE start_time BETWEEN '2018-07-20 00:00:00' AND '2018-07-21 00:00:00';"

         tree         |    field    |                      description
+---------------------+-------------+-------------------------------------------------------+
  group               |             |
   │                  | aggregate 0 | count(DISTINCT id)
   │                  | scalar      |
   └── render         |             |
        └── join      |             |
             │        | type        | inner
             │        | equality    | (id) = (rider_id)
             ├── scan |             |
             │        | table       | users@users_name_idx
             │        | spans       | ALL
             └── scan |             |
                      | table       | rides@rides_start_time_idx
                      | spans       | /2018-07-20T00:00:00Z-/2018-07-21T00:00:00.000000001Z
(13 rows)

Notice that CockroachDB now starts by using rides@rides_start_time_idx secondary index to retrieve the relevant rides without needing to scan the full rides table.

Let's check the ranges for the new index:

$ cockroach sql \
--certs-dir=certs \
--host=<address of any node> \
--database=movr \
--execute="SHOW EXPERIMENTAL_RANGES FROM INDEX rides@rides_start_time_idx;"

                                         start_key                                         |                                         end_key                                          | range_id | replicas | lease_holder
+------------------------------------------------------------------------------------------+------------------------------------------------------------------------------------------+----------+----------+--------------+
  NULL                                                                                     | /2018-07-11T01:37:36.138325Z/"new york"/"\xd4\xe3\u007f\xbc2\xc0Mv\x81B\xd6\xc7٘\x9f\xe6" |       45 | {1,2,3}  |            2
  /2018-07-11T01:37:36.138325Z/"new york"/"\xd4\xe3\u007f\xbc2\xc0Mv\x81B\xd6\xc7٘\x9f\xe6" | NULL                                                                                     |       50 | {1,2,3}  |            2
(2 rows)

This tells us that the index is stored in 2 ranges, with the leaseholders for both of them on node 2. Based on the output of SHOW EXPERIMENTAL_RANGES FROM TABLE users that we saw earlier, we already know that the leaseholder for the users table is on node 2.

Using IN (list) with a subquery

Now let's say you want to get the latest ride of each of the 5 most used vehicles. To do this, you might think to use a subquery to get the IDs of the 5 most frequent vehicles from the rides table, passing the results into the IN list of another query to get the most recent ride of each of the 5 vehicles:

$ ./tuning-secure.py \
--host=<address of any node> \
--statement="SELECT vehicle_id, max(end_time) \
FROM rides \
WHERE vehicle_id IN ( \
    SELECT vehicle_id \
    FROM rides \
    GROUP BY vehicle_id \
    ORDER BY count(*) DESC \
    LIMIT 5 \
) \
GROUP BY vehicle_id" \
--repeat=20 \
--times

Result:
['vehicle_id', 'max']
['3c950d36-c2b8-48d0-87d3-e0d6f570af62', '2018-08-02 03:06:31.293184']
['0962cdca-9d85-457c-9616-cc2ae2d32008', '2018-08-02 03:01:25.414512']
['78fdd6f8-c6a1-42df-a89f-cd65b7bb8be9', '2018-08-02 02:47:43.755989']
['c6541da5-9858-4e3f-9b49-992e206d2c50', '2018-08-02 02:14:50.543760']
['35752c4c-b878-4436-8330-8d7246406a55', '2018-08-02 03:08:49.823209']
Times (milliseconds):
[3012.6540660858154, 2456.5110206604004, 2482.675075531006, 2488.3930683135986, 2474.393129348755, 2494.3790435791016, 2504.063129425049, 2491.326093673706, 2507.4589252471924, 2482.077121734619, 2495.9230422973633, 2497.60103225708, 2478.4271717071533, 2496.574878692627, 2506.395101547241, 2468.4300422668457, 2476.508140563965, 2497.958183288574, 2480.7958602905273, 2484.0168952941895]
Median time (milliseconds):
2489.85958099

However, as you can see, this query is slow because, currently, when the WHERE condition of a query comes from the result of a subquery, CockroachDB scans the entire table, even if there is an available index. Use EXPLAIN to see this in more detail:

$ cockroach sql \
--certs-dir=certs \
--host=<address of any node> \
--database=movr \
--execute="EXPLAIN SELECT vehicle_id, max(end_time) \
FROM rides \
WHERE vehicle_id IN ( \
    SELECT vehicle_id \
    FROM rides \
    GROUP BY vehicle_id \
    ORDER BY count(*) DESC \
    LIMIT 5 \
) \
GROUP BY vehicle_id;"

              tree              |    field    |                     description
+-------------------------------+-------------+-----------------------------------------------------+
  group                         |             |
   │                            | aggregate 0 | vehicle_id
   │                            | aggregate 1 | max(end_time)
   │                            | group by    | @1
   └── join                     |             |
        │                       | type        | semi
        │                       | equality    | (vehicle_id) = (vehicle_id)
        ├── scan                |             |
        │                       | table       | rides@primary
        │                       | spans       | ALL
        └── limit               |             |
             └── sort           |             |
                  │             | order       | -agg0
                  └── group     |             |
                       │        | aggregate 0 | vehicle_id
                       │        | aggregate 1 | count_rows()
                       │        | group by    | @1
                       └── scan |             |
                                | table       | rides@rides_auto_index_fk_vehicle_city_ref_vehicles
                                | spans       | ALL
(20 rows)

This is a complex query plan, but the important thing to note is the full table scan of rides@primary above the subquery. This shows you that, after the subquery returns the IDs of the top 5 vehicles, CockroachDB scans the entire primary index to find the rows with max(end_time) for each vehicle_id, although you might expect CockroachDB to more efficiently use the secondary index on vehicle_id (CockroachDB is working to remove this limitation in a future version).

Using IN (list) with explicit values

Because CockroachDB will not use an available secondary index when using IN (list) with a subquery, it's much more performant to have your application first select the top 5 vehicles:

$ ./tuning-secure.py \
--host=<address of any node> \
--statement="SELECT vehicle_id \
FROM rides \
GROUP BY vehicle_id \
ORDER BY count(*) DESC \
LIMIT 5" \
--repeat=20 \
--times

Result:
['vehicle_id']
['35752c4c-b878-4436-8330-8d7246406a55']
['0962cdca-9d85-457c-9616-cc2ae2d32008']
['c6541da5-9858-4e3f-9b49-992e206d2c50']
['78fdd6f8-c6a1-42df-a89f-cd65b7bb8be9']
['3c950d36-c2b8-48d0-87d3-e0d6f570af62']
Times (milliseconds):
[1049.2329597473145, 1038.0151271820068, 1037.7991199493408, 1036.5591049194336, 1037.7249717712402, 1040.544033050537, 1022.7780342102051, 1056.9651126861572, 1054.3549060821533, 1042.3550605773926, 1042.68217086792, 1031.7370891571045, 1051.880121231079, 1035.8471870422363, 1035.2818965911865, 1035.607099533081, 1040.0230884552002, 1048.8879680633545, 1056.014060974121, 1036.1089706420898]
Median time (milliseconds):
1039.01910782

And then put the results into the IN list to get the most recent rides of the vehicles:

$ ./tuning-secure.py \
--host=<address of any node> \
--statement="SELECT vehicle_id, max(end_time) \
FROM rides \
WHERE vehicle_id IN ( \
  '35752c4c-b878-4436-8330-8d7246406a55', \
  '0962cdca-9d85-457c-9616-cc2ae2d32008', \
  'c6541da5-9858-4e3f-9b49-992e206d2c50', \
  '78fdd6f8-c6a1-42df-a89f-cd65b7bb8be9', \
  '3c950d36-c2b8-48d0-87d3-e0d6f570af62' \
) \
GROUP BY vehicle_id;" \
--repeat=20 \
--times

Result:
['vehicle_id', 'max']
['35752c4c-b878-4436-8330-8d7246406a55', '2018-08-02 03:08:49.823209']
['0962cdca-9d85-457c-9616-cc2ae2d32008', '2018-08-02 03:01:25.414512']
['3c950d36-c2b8-48d0-87d3-e0d6f570af62', '2018-08-02 03:06:31.293184']
['78fdd6f8-c6a1-42df-a89f-cd65b7bb8be9', '2018-08-02 02:47:43.755989']
['c6541da5-9858-4e3f-9b49-992e206d2c50', '2018-08-02 02:14:50.543760']
Times (milliseconds):
[1165.5981540679932, 1135.9851360321045, 1201.0550498962402, 1135.0820064544678, 1195.7061290740967, 1132.0109367370605, 1134.9878311157227, 1175.88210105896, 1174.0548610687256, 1152.566909790039, 1164.9351119995117, 1175.5108833312988, 1161.651849746704, 1195.3318119049072, 1162.4629497528076, 1156.1191082000732, 1127.0110607147217, 1165.4651165008545, 1159.6789360046387, 1190.3491020202637]
Median time (milliseconds):
1163.69903088

This approach reduced the query time from 2489.85ms (query with subquery) to 2202.70ms (2 distinct queries).

Step 7. Test/tune write performance

• Bulk inserting into an existing table
• Minimizing unused indexes
• Retrieving the ID of a newly inserted row

Bulk inserting into an existing table

Moving on to writes, let's imagine that you have a batch of 100 new users to insert into the users table. The most obvious approach is to insert each row using 100 separate INSERT statements:

Note:

For the purpose of demonstration, the command below inserts the same user 100 times, each time with a different unique ID. Note also that you're now adding the —cumulative flag to print the total time across all 100 inserts.

$ ./tuning-secure.py \
--host=<address of any node> \
--statement="INSERT INTO users VALUES (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347')" \
--repeat=100 \
--times \
--cumulative

Times (milliseconds):
[10.773181915283203, 12.186050415039062, 9.711980819702148, 9.730815887451172, 10.200977325439453, 9.32002067565918, 9.002923965454102, 9.426116943359375, 9.312152862548828, 8.329153060913086, 9.626150131225586, 8.965015411376953, 9.562969207763672, 9.305000305175781, 9.34910774230957, 7.394075393676758, 9.3231201171875, 9.066104888916016, 8.419036865234375, 9.158134460449219, 9.278059005737305, 8.022069931030273, 8.542060852050781, 9.237051010131836, 8.165121078491211, 8.094072341918945, 8.025884628295898, 8.04591178894043, 9.728193283081055, 8.485078811645508, 7.967948913574219, 9.319067001342773, 8.099079132080078, 9.041070938110352, 10.046005249023438, 10.684013366699219, 9.672880172729492, 8.129119873046875, 8.10098648071289, 7.884979248046875, 9.484052658081055, 8.594036102294922, 9.479045867919922, 9.239912033081055, 9.16600227355957, 9.155988693237305, 9.392976760864258, 11.08694076538086, 9.402990341186523, 8.034944534301758, 8.053064346313477, 8.03995132446289, 8.891820907592773, 8.054971694946289, 8.903980255126953, 9.057998657226562, 9.713888168334961, 7.99107551574707, 8.114814758300781, 8.677959442138672, 11.178970336914062, 9.272098541259766, 9.281158447265625, 8.177995681762695, 9.47880744934082, 10.025978088378906, 8.352041244506836, 8.320808410644531, 10.892868041992188, 8.227825164794922, 8.220911026000977, 9.625911712646484, 10.272026062011719, 8.116960525512695, 10.786771774291992, 9.073972702026367, 9.686946868896484, 9.903192520141602, 9.887933731079102, 9.399890899658203, 9.413003921508789, 8.594036102294922, 8.433103561401367, 9.271860122680664, 8.529901504516602, 9.474992752075195, 9.005069732666016, 9.341001510620117, 9.388923645019531, 9.775876998901367, 8.558988571166992, 9.613990783691406, 8.897066116333008, 8.642911911010742, 9.527206420898438, 8.274078369140625, 9.073972702026367, 9.637832641601562, 8.516788482666016, 9.564876556396484]
Median time (milliseconds):
9.20152664185
Cumulative time (milliseconds):
910.985708237

The 100 inserts took 910.98ms to complete, which isn't bad. However, it's significantly faster to use a single INSERT statement with 100 comma-separated VALUES clauses:

$ ./tuning-secure.py \
--host=<address of any node> \
--statement="INSERT INTO users VALUES \
(gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), \
(gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), \
(gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), \
(gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), \
(gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), \
(gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), \
(gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), \
(gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), \
(gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), \
(gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347'), (gen_random_uuid(), 'new york', 'Max Roach', '411 Drum Street', '173635282937347')" \
--repeat=1 \
--cumulative

Median time (milliseconds):
15.4001712799
Cumulative time (milliseconds):
15.4001712799

As you can see, this multi-row INSERT technique reduced the total time for 100 inserts from 910.98ms to 15.40ms. It's useful to note that this technique is equally effective for UPSERT and DELETE statements as well.

Minimizing unused indexes

Earlier, we saw how important secondary indexes are for read performance. For writes, however, it's important to recognized the overhead that they create.

Let's consider the users table:

$ cockroach sql \
--certs-dir=certs \
--host=<address of any node> \
--database=movr \
--execute="SHOW INDEXES FROM users;"

  table_name |   index_name   | non_unique | seq_in_index | column_name | direction | storing | implicit
+------------+----------------+------------+--------------+-------------+-----------+---------+----------+
  users      | primary        |   false    |            1 | city        | ASC       |  false  |  false
  users      | primary        |   false    |            2 | id          | ASC       |  false  |  false
  users      | users_name_idx |    true    |            1 | name        | ASC       |  false  |  false
  users      | users_name_idx |    true    |            2 | credit_card | N/A       |  true   |  false
  users      | users_name_idx |    true    |            3 | city        | ASC       |  false  |   true
  users      | users_name_idx |    true    |            4 | id          | ASC       |  false  |   true
(6 rows)

This table has the primary index (the full table) and a secondary index on name that is also storing credit_card. This means that whenever a row is inserted, or whenever name, credit_card, city, or id are modified in existing rows, both indexes are updated.

To make this more concrete, let's count how many rows have a name that starts with C and then update those rows to all have the same name:

$ ./tuning-secure.py \
--host=<address of any node> \
--statement="SELECT count(*) \
FROM users \
WHERE name LIKE 'C%'" \
--repeat=1

['count']
['179']
Median time (milliseconds):
2.52604484558

$ ./tuning-secure.py \
--host=<address of any node> \
--statement="UPDATE users \
SET name = 'Carl Kimball' \
WHERE name LIKE 'C%'" \
--repeat=1

Median time (milliseconds):
52.2060394287

Because name is in both the primary and users_name_idx indexes, for each of the 168 rows, 2 keys were updated.

Now, assuming that the users_name_idx index is no longer needed, lets drop the index and execute an equivalent query:

$ cockroach sql \
--certs-dir=certs \
--host=<address of any node> \
--database=movr \
--execute="DROP INDEX users_name_idx;"

$ ./tuning-secure.py \
--host=<address of any node> \
--statement="UPDATE users \
SET name = 'Peedie Hirata' \
WHERE name = 'Carl Kimball'" \
--repeat=1

Median time (milliseconds):
22.7289199829

Before, when both the primary and secondary indexes needed to be updated, the updates took 52.20ms. Now, after dropping the secondary index, an equivalent update took only 22.72ms.

Retrieving the ID of a newly inserted row

Now let's focus on the common case of inserting a row into a table and then retrieving the ID of the new row to do some follow-up work. One approach is to execute two statements, an INSERT to insert the row and then a SELECT to get the new ID:

$ ./tuning-secure.py \
--host=<address of any node> \
--statement="INSERT INTO users VALUES (gen_random_uuid(), 'new york', 'Toni Brooks', '800 Camden Lane, Brooklyn, NY 11218', '98244843845134960')" \
--repeat=1

Median time (milliseconds):
10.4398727417

$ ./tuning-secure.py \
--host=<address of any node> \
--statement="SELECT id FROM users WHERE name = 'Toni Brooks'" \
--repeat=1

Result:
['id']
['ae563e17-ad59-4307-a99e-191e682b4278']
Median time (milliseconds):
5.53798675537

Combined, these statements are relatively fast, at 15.96ms, but an even more performant approach is to append RETURNING id to the end of the INSERT:

$ ./tuning-secure.py \
--host=<address of any node> \
--statement="INSERT INTO users VALUES (gen_random_uuid(), 'new york', 'Brian Brooks', '800 Camden Lane, Brooklyn, NY 11218', '98244843845134960') \
RETURNING id" \
--repeat=1

Result:
['id']
['3d16500e-cb2e-462e-9c83-db0965d6deaf']
Median time (milliseconds):
9.48596000671

At just 9.48ms, this approach is faster due to the write and read executing in one instead of two client-server roundtrips. Note also that, as discussed earlier, if the leaseholder for the table happens to be on a different node than the query is running against, that introduces additional network hops and latency.

Multi-region deployment

Given that Movr is active on both US coasts, you'll now scale the cluster into two new regions, us-west1-a and us-west2-a, each with 3 nodes and an extra instance for simulating regional client traffic.

Step 8. Create more instances

• Create 6 more instances, 3 in the us-west1-a zone (Oregon), and 3 in the us-west2-a zone (Los Angeles). While creating each instance:

  • Use the n1-standard-4 machine type (4 vCPUs, 15 GB memory).
  • Use the Ubuntu 16.04 OS image.
  • Create and mount a local SSD.
  • To apply the Web UI firewall rule you created earlier, click Management, disk, networking, SSH keys, select the Networking tab, and then enter cockroachdb in the Network tags field.

• Note the internal and external IP addresses of each n1-standard-4 instance. You'll need these addresses when generating security certificates and when starting the CockroachDB nodes.

• Create an additional instance in the us-west1-a and us-west2-a zones. These can be smaller, such as n1-standard-1.

Step 9. Scale the cluster

Generate security certificates

• On your local machine, where you generated certificates for your first nodes, create the certificate and key for one of the new nodes, issued to all common names you might use to refer to the node:

$ cockroach cert create-node \
<node internal IP address> \
<node external IP address> \
<node hostname>  \
<other common names for node> \
localhost \
127.0.0.1 \
--certs-dir=certs \
--ca-key=my-safe-directory/ca.key

• Upload certificates to the instance:

# Create the certs directory:
$ ssh <username>@<node address> "mkdir certs"

# Upload the CA certificate and node certificate and key:
$ scp certs/ca.crt \
certs/node.crt \
certs/node.key \
<username>@<node address>:~/certs

• Delete the local copy of the node certificate and key:

$ rm certs/node.crt certs/node.key

Note:

This is necessary because the certificates and keys for additional nodes will also be named node.crt and node.key. As an alternative to deleting these files, you can run the next cockroach cert create-node commands with the —overwrite flag.

• Repeat steps 1 - 3 for each new node.

• Upload the client certificates you created earlier to the additional instances in the us-west1-a and us-west2-a zones for your client application workload:

# Create the certs directory:
$ ssh <username>@<instance address> "mkdir certs"

# Upload the CA certificate and client certificate and key:
$ scp certs/ca.crt \
certs/client.root.crt \
certs/client.root.key \
<username>@<instance address>:~/certs

Start the new nodes

• SSH to one of the n1-standard-4 instances in the us-west1-a zone.

• Download the CockroachDB archive for Linux, extract the binary, and copy it into the PATH:

$ wget -qO- https://binaries.cockroachdb.com/cockroach-v19.1.0.linux-amd64.tgz \
| tar  xvz

$ sudo cp -i cockroach-v19.1.0.linux-amd64/cockroach /usr/local/bin

• Run the cockroach start command:

$ cockroach start \
--certs-dir=certs \
--advertise-host=<node internal address> \
--join=<same as earlier> \
--locality=cloud=gce,region=us-west1,zone=us-west1-a \
--cache=.25 \
--max-sql-memory=.25 \
--background

• Repeat steps 1 - 3 for the other two n1-standard-4 instances in the us-west1-a zone.

• SSH to one of the n1-standard-4 instances in the us-west2-a zone.

• Download the CockroachDB archive for Linux, extract the binary, and copy it into the PATH:

$ wget -qO- https://binaries.cockroachdb.com/cockroach-v19.1.0.linux-amd64.tgz \
| tar  xvz

$ sudo cp -i cockroach-v19.1.0.linux-amd64/cockroach /usr/local/bin

• Run the cockroach start command:

$ cockroach start \
--certs-dir=certs \
--advertise-host=<node1 internal address> \
--join=<same as earlier> \
--locality=cloud=gce,region=us-west2,zone=us-west2-a \
--cache=.25 \
--max-sql-memory=.25 \
--background

• Repeat steps 5 - 7 for the other two n1-standard-4 instances in the us-west2-a zone.

Step 10. Install the Python client

In each of the new zones, SSH to the instance not running a CockroachDB node, and install the Python client as described in step 5 above.

Step 11. Check rebalancing

Since you started each node with the —locality flag set to its GCE zone, over the next minutes, CockroachDB will rebalance data evenly across the zones.

To check this, access the Web UI on any node at <node address>:8080 and look at the Node List. You'll see that the range count is more or less even across all nodes:

For reference, here's how the nodes map to zones:

To verify even balancing at range level, SSH to one of the instances not running CockroachDB and run the SHOW EXPERIMENTAL_RANGES statement:

$ cockroach sql \
--certs-dir=certs \
--host=<address of any node> \
--database=movr \
--execute="SHOW EXPERIMENTAL_RANGES FROM TABLE vehicles;"

  start_key | end_key | range_id | replicas | lease_holder
+-----------+---------+----------+----------+--------------+
  NULL      | NULL    |       33 | {3,4,7}  |            7
(1 row)

In this case, we can see that, for the single range containing vehicles data, one replica is in each zone, and the leaseholder is in the us-west2-a zone.

Step 12. Test performance

In general, all of the tuning techniques featured in the single-region scenario above still apply in a multi-region deployment. However, the fact that data and leaseholders are spread across the US means greater latencies in many cases.

Reads

For example, imagine we are a Movr administrator in New York, and we want to get the IDs and descriptions of all New York-based bikes that are currently in use:

• SSH to the instance in us-east1-b with the Python client.

• Query for the data:

$ ./tuning-secure.py \
--host=<address of a node in us-east1-b> \
--statement="SELECT id, ext FROM vehicles \
WHERE city = 'new york' \
    AND type = 'bike' \
    AND status = 'in_use'" \
--repeat=50 \
--times

Result:
['id', 'ext']
['0068ee24-2dfb-437d-9a5d-22bb742d519e', "{u'color': u'green', u'brand': u'Kona'}"]
['01b80764-283b-4232-8961-a8d6a4121a08', "{u'color': u'green', u'brand': u'Pinarello'}"]
['02a39628-a911-4450-b8c0-237865546f7f', "{u'color': u'black', u'brand': u'Schwinn'}"]
['02eb2a12-f465-4575-85f8-a4b77be14c54', "{u'color': u'black', u'brand': u'Pinarello'}"]
['02f2fcc3-fea6-4849-a3a0-dc60480fa6c2', "{u'color': u'red', u'brand': u'FujiCervelo'}"]
['034d42cf-741f-428c-bbbb-e31820c68588', "{u'color': u'yellow', u'brand': u'Santa Cruz'}"]
...
Times (milliseconds):
[933.8209629058838, 72.02410697937012, 72.45206832885742, 72.39294052124023, 72.8158950805664, 72.07584381103516, 72.21412658691406, 71.96712493896484, 71.75517082214355, 72.16811180114746, 71.78592681884766, 72.91603088378906, 71.91109657287598, 71.4719295501709, 72.40676879882812, 71.8080997467041, 71.84004783630371, 71.98500633239746, 72.40891456604004, 73.75001907348633, 71.45905494689941, 71.53081893920898, 71.46596908569336, 72.07608222961426, 71.94995880126953, 71.41804695129395, 71.29096984863281, 72.11899757385254, 71.63381576538086, 71.3050365447998, 71.83194160461426, 71.20394706726074, 70.9981918334961, 72.79205322265625, 72.63493537902832, 72.15285301208496, 71.8698501586914, 72.30591773986816, 71.53582572937012, 72.69001007080078, 72.03006744384766, 72.56317138671875, 71.61688804626465, 72.17121124267578, 70.20092010498047, 72.12018966674805, 73.34589958190918, 73.01592826843262, 71.49410247802734, 72.19099998474121]
Median time (milliseconds):
72.0270872116

As we saw earlier, the leaseholder for the vehicles table is in us-west2-a (Los Angeles), so our query had to go from the gateway node in us-east1-b all the way to the west coast and then back again before returning data to the client.

For contrast, imagine we are now a Movr administrator in Los Angeles, and we want to get the IDs and descriptions of all Los Angeles-based bikes that are currently in use:

• SSH to the instance in us-west2-a with the Python client.

• Query for the data:

$ ./tuning-secure.py \
--host=<address of a node in us-west2-a> \
--statement="SELECT id, ext FROM vehicles \
WHERE city = 'los angeles' \
    AND type = 'bike' \
    AND status = 'in_use'" \
--repeat=50 \
--times

Result:
['id', 'ext']
['00078349-94d4-43e6-92be-8b0d1ac7ee9f', "{u'color': u'blue', u'brand': u'Merida'}"]
['003f84c4-fa14-47b2-92d4-35a3dddd2d75', "{u'color': u'red', u'brand': u'Kona'}"]
['0107a133-7762-4392-b1d9-496eb30ee5f9', "{u'color': u'yellow', u'brand': u'Kona'}"]
['0144498b-4c4f-4036-8465-93a6bea502a3', "{u'color': u'blue', u'brand': u'Pinarello'}"]
['01476004-fb10-4201-9e56-aadeb427f98a', "{u'color': u'black', u'brand': u'Merida'}"]
Times (milliseconds):
[782.6759815216064, 8.564949035644531, 8.226156234741211, 7.949113845825195, 7.86590576171875, 7.842063903808594, 7.674932479858398, 7.555961608886719, 7.642984390258789, 8.024930953979492, 7.717132568359375, 8.46409797668457, 7.520914077758789, 7.6541900634765625, 7.458925247192383, 7.671833038330078, 7.740020751953125, 7.771015167236328, 7.598161697387695, 8.411169052124023, 7.408857345581055, 7.469892501831055, 7.524967193603516, 7.764101028442383, 7.750988006591797, 7.2460174560546875, 6.927967071533203, 7.822990417480469, 7.27391242980957, 7.730960845947266, 7.4710845947265625, 7.4310302734375, 7.33494758605957, 7.455110549926758, 7.021188735961914, 7.083892822265625, 7.812976837158203, 7.625102996826172, 7.447957992553711, 7.179021835327148, 7.504940032958984, 7.224082946777344, 7.257938385009766, 7.714986801147461, 7.4939727783203125, 7.6160430908203125, 7.578849792480469, 7.890939712524414, 7.546901702880859, 7.411956787109375]
Median time (milliseconds):
7.6071023941