Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
It's 10pm: Do You Know Where Your Writes Are?
Search
Jeremy Mikola
October 12, 2017
Programming
0
210
It's 10pm: Do You Know Where Your Writes Are?
Presented October 12, 2017 at MongoDB.local San Francisco.
Jeremy Mikola
October 12, 2017
Tweet
Share
More Decks by Jeremy Mikola
See All by Jeremy Mikola
PHP Internals for the Inquisitive Developer
jmikola
1
650
Bulletproof MongoDB
jmikola
0
480
Zero to Sixty with MongoDB
jmikola
3
1.1k
DOs and DON’Ts of MongoDB
jmikola
13
3.1k
Five Years of Beta
jmikola
0
150
Rethinking Extension Development for PHP and HHVM
jmikola
2
830
What's New in MongoDB 3.2
jmikola
0
110
Async PHP with React
jmikola
28
11k
NoSQL Lightning Talks (MongoDB, Cassandra, MySQL)
jmikola
1
250
Other Decks in Programming
See All in Programming
Kiroの仕様駆動開発から見えてきたAIコーディングとの正しい付き合い方
clshinji
1
190
Swift Updates - Learn Languages 2025
koher
1
250
フロントエンドのmonorepo化と責務分離のリアーキテクト
kajitack
2
160
Azure SRE Agentで運用は楽になるのか?
kkamegawa
0
1.1k
Kiroで始めるAI-DLC
kaonash
2
510
UbieのAIパートナーを支えるコンテキストエンジニアリング実践
syucream
2
800
Updates on MLS on Ruby (and maybe more)
sylph01
1
170
もうちょっといいRubyプロファイラを作りたい (2025)
osyoyu
0
230
250830 IaCの選定~AWS SAMのLambdaをECSに乗り換えたときの備忘録~
east_takumi
0
360
AIでLINEスタンプを作ってみた
eycjur
1
220
複雑なドメインに挑む.pdf
yukisakai1225
5
910
Claude Codeで挑むOSSコントリビュート
eycjur
0
190
Featured
See All Featured
How to train your dragon (web standard)
notwaldorf
96
6.2k
BBQ
matthewcrist
89
9.8k
Cheating the UX When There Is Nothing More to Optimize - PixelPioneers
stephaniewalter
284
13k
The World Runs on Bad Software
bkeepers
PRO
70
11k
The Pragmatic Product Professional
lauravandoore
36
6.8k
How STYLIGHT went responsive
nonsquared
100
5.8k
YesSQL, Process and Tooling at Scale
rocio
173
14k
Responsive Adventures: Dirty Tricks From The Dark Corners of Front-End
smashingmag
252
21k
The Art of Delivering Value - GDevCon NA Keynote
reverentgeek
15
1.6k
How to Ace a Technical Interview
jacobian
279
23k
JavaScript: Past, Present, and Future - NDC Porto 2020
reverentgeek
51
5.6k
Reflections from 52 weeks, 52 projects
jeffersonlam
351
21k
Transcript
It’s 10PM: Do you know where your writes are? Jeremy
Mikola jmikola
It’s 11AM: Do you know where your writes are? Jeremy
Mikola jmikola
On the roadmap Retryable writes Zombie cursor cleanup Cluster-wide killOp
Retryable Writes
Retryable Writes
Retryable Writes
You’re updating a document
You’re updating a document db.coll.updateOne( { _id: 16 }, {
$inc: { count: 1 }} );
Murphy’s Law kicks in
Is it safe to retry the update?
Did our message never make it to the server?
Did we lose the server’s reply?
Did something else happen?
Did something else happen?
Did something else happen?
Did something else happen?
Did something else happen?
There’s no way to retrieve an operation’s state
Let’s review some best practices How To Write Resilient MongoDB
Applications
This was our update… db.coll.updateOne( { _id: 16 }, {
$inc: { count: 1 }} );
Errors and retry strategies Transient network error Persistent outage Command
error Never retry Always retry Retry once
Errors and retry strategies Transient network error Persistent outage Command
error Never retry May undercount Always retry Retry once
Errors and retry strategies Transient network error Persistent outage Command
error Never retry May undercount OK Always retry Retry once
Errors and retry strategies Transient network error Persistent outage Command
error Never retry May undercount OK OK Always retry Retry once
Errors and retry strategies Transient network error Persistent outage Command
error Never retry May undercount OK OK Always retry May overcount Retry once
Errors and retry strategies Transient network error Persistent outage Command
error Never retry May undercount OK OK Always retry May overcount Wastes time Retry once
Errors and retry strategies Transient network error Persistent outage Command
error Never retry May undercount OK OK Always retry May overcount Wastes time Wastes time Retry once
Errors and retry strategies Transient network error Persistent outage Command
error Never retry May undercount OK OK Always retry May overcount Wastes time Wastes time Retry once May overcount
Errors and retry strategies Transient network error Persistent outage Command
error Never retry May undercount OK OK Always retry May overcount Wastes time Wastes time Retry once May overcount OK
Errors and retry strategies Transient network error Persistent outage Command
error Never retry May undercount OK OK Always retry May overcount Wastes time Wastes time Retry once May overcount OK OK
There’s no good solution for transient network errors
We can safely retry idempotent operations Transient network error Persistent
outage Command error Retry once
We can safely retry idempotent operations Transient network error Persistent
outage Command error Retry once OK
We can safely retry idempotent operations Transient network error Persistent
outage Command error Retry once OK OK
We can safely retry idempotent operations Transient network error Persistent
outage Command error Retry once OK OK OK
Safe-to-retry inserts db.coll.insertOne( { _id: 18, name: "Alice" } );
Safe-to-retry deletes db.coll.deleteOne( { _id: 20 } ); db.coll.deleteMany( {
status: "inactive" } );
Safe-to-retry updates db.coll.updateOne( { _id: 22 }, { $set: {
status: "active" }} );
Why can’t we retrieve an operation’s state?
In MongoDB 3.4, state is tied to connection objects
MongoDB 3.6 introduces logical sessions
MongoDB 3.6 introduces logical sessions Sessions allow us to maintain
cluster-wide state about the user and their operations.
MongoDB 3.6 introduces logical sessions Sessions allow us to maintain
cluster-wide state about the user and their operations. Sessions are not tied to connections.
Retrying writes with a session
Retrying writes with a session
Retrying writes with a session
Retrying writes with a session update
Retrying writes with a session
Retrying writes with a session update
Retrying writes with a session
We can trust the server to Do the Right Thing™
We can trust the server to Do the Right Thing™
If the write already executed, return the result we missed.
We can trust the server to Do the Right Thing™
If the write already executed, return the result we missed. If the write never executed, do it now and return its result.
Sessions are cluster-wide
Sessions are cluster-wide update
Sessions are cluster-wide update
Sessions are cluster-wide
Sessions are cluster-wide
Sessions are cluster-wide update
Taking advantage of retryable writes ?retryWrites=true mongodb://…
One down, two to go Retryable writes Zombie cursor cleanup
Cluster-wide killOp
Zombie Cursor Cleanup
Zombie Cursor Cleanup
You’re running a long query
You’re running a long query cursor = db.coll.find(); cursor.forEach(function() {
// lengthy processing… });
You’re running a long query cursor = db.coll.find(); cursor.forEach(function() {
// lengthy processing… });
Cursors have a timeout
Cursors have a timeout A er 10 minutes, the server
will close a cursor due to inactivity.
Cursors have a timeout A er 10 minutes, the server
will close a cursor due to inactivity. Issuing a getMore resets the clock.
Disabling cursor timeouts cursor = db.coll.find( { }, { noCursorTimeout:
true } ); cursor.forEach(function() { // lengthy processing…
Disabling cursor timeouts cursor = db.coll.find( { }, { noCursorTimeout:
true } ); cursor.forEach(function() { // lengthy processing…
Executing our long query
Executing our long query find
Executing our long query
Executing our long query getMore
Executing our long query
Executing our long query getMore
Executing our long query
Executing our long query getMore
Executing our long query getMore
Executing our long query getMore
Executing our long query getMore
Executing our long query getMore
Executing our long query getMore
Executing our long query getMore
None
None
None
None
A zombie cursor is born > db.serverStatus() { "metrics": {
"cursor": { "open": { "noTimeout": 1, "total": 1
What happened last night?
What happened last night? (from the server’s POV)
What happened last night? (from the server’s POV)
What happened last night? (from the server’s POV) find
What happened last night? (from the server’s POV)
What happened last night? (from the server’s POV) getMore
What happened last night? (from the server’s POV)
What happened last night? (from the server’s POV) getMore
What happened last night? (from the server’s POV)
What happened last night? (from the server’s POV)
What happened last night? (from the server’s POV)
Avoiding zombie cursors with logical sessions
Avoiding zombie cursors with logical sessions Sessions also have a
timeout.
Avoiding zombie cursors with logical sessions Sessions also have a
timeout. We can associate queries with a session
Querying with a session session = client.startSession(); cursor = db.coll.find(
{ }, { session: session } );
Executing our long query
Executing our long query
Executing our long query
Executing our long query find
Executing our long query
Executing our long query getMore
Executing our long query
Executing our long query getMore
Executing our long query
Executing our long query
Executing our long query session expires
Executing our long query
Did we just punt on the timeout issue?
Session timeouts are non-negotiable
Session timeouts are non-negotiable Idle sessions will expire.
Session timeouts are non-negotiable Idle sessions will expire. Any operation
using the session resets the clock.
Two down, one to go Retryable writes Zombie cursor cleanup
Cluster-wide killOp
Cluster-wide killOp
Cluster-wide killOp
You’re running an operation that may never complete
You’re running an operation that may never complete cursor =
db.coll.find( { … } // table scans for days );
You’ve made a terrible mistake
Step 1: Find the operation ID > db.currentOp() { "inprog"
: [ { "desc" : "conn2", "threadId" : "140181791471360", "connectionId" : 2, "client" : "127.0.0.1:49456", "appName" : "MongoDB Shell", "active" : true, "opid" : 132921,
Step 2: Kill the operation ID > db.killOp(132921) { "info":
"attempting to kill op", "ok": 1 }
Lather, rinse, repeat
Lather, rinse, repeat > connect("mongodb://shard-2.example.com")
Lather, rinse, repeat > connect("mongodb://shard-2.example.com") > db.currentOp() { "inprog" :
[ // … ] }
Lather, rinse, repeat > connect("mongodb://shard-2.example.com") > db.currentOp() { "inprog" :
[ // … ] } > db.killOp(…)
Lather, rinse, repeat > connect("mongodb://shard-2.example.com") > db.currentOp() { "inprog" :
[ // … ] } > db.killOp(…)
How did this happen? mongos shard 1 shard 2 shard
3
How did this happen? mongos shard 1 shard 2 shard
3
How did this happen? mongos shard 1 shard 2 shard
3
Cluster-wide killOp with logical sessions
Cluster-wide killOp with logical sessions Any operation may be associated
with a session.
Cluster-wide killOp with logical sessions Any operation may be associated
with a session. Terminating a session will end all of its associated operations.
Terminating a session session = client.startSession(); cursor = db.coll.find( {
… }, // table scans for days { session: session } );
Querying with sessions mongos shard 1 shard 2 shard 3
Querying with sessions mongos shard 1 shard 2 shard 3
Querying with sessions mongos shard 1 shard 2 shard 3
Querying with sessions mongos shard 1 shard 2 shard 3
Querying with sessions mongos shard 1 shard 2 shard 3
Querying with sessions mongos shard 1 shard 2 shard 3
Querying with sessions mongos shard 1 shard 2 shard 3
Querying with sessions mongos shard 1 shard 2 shard 3
Querying with sessions mongos shard 1 shard 2 shard 3
Querying with sessions mongos shard 1 shard 2 shard 3
That’s a wrap Retryable writes Zombie cursor cleanup Cluster-wide killOp
One last point
Resilence is primarily the driver’s domain
Resilence is primarily the driver’s domain Server discovery and monitoring
Resilence is primarily the driver’s domain Server discovery and monitoring
Elections and failover recovery
Resilence is primarily the driver’s domain Server discovery and monitoring
Elections and failover recovery Load-balancing mongos connections
Resilence is primarily the driver’s domain Server discovery and monitoring
Elections and failover recovery Load-balancing mongos connections Routing queries by read preference
Addressing resilence on the server-side
Addressing resilence on the server-side Tracking operation state
Addressing resilence on the server-side Tracking operation state Cluster-wide sessions
Providing a relatively easy upgrade path
Providing a relatively easy upgrade path No need to rewrite
applications
Providing a relatively easy upgrade path No need to rewrite
applications Opting in to retryable writes
Providing a relatively easy upgrade path No need to rewrite
applications Opting in to retryable writes New API for client session objects
Providing a relatively easy upgrade path No need to rewrite
applications Opting in to retryable writes New API for client session objects Pass session option as needed
Inside the spec process mongodb/specifications
Inside the spec process /sessions mongodb/specifications
Inside the spec process /sessions /retryable-writes mongodb/specifications
Inside the spec process /sessions /retryable-writes /causal-consistency mongodb/specifications
Inside the spec process /sessions /retryable-writes /causal-consistency /retryable-reads mongodb/specifications
In the meantime… How To Write Resilient MongoDB Applications
Thanks!