Etsy Activity Feeds Architecture

Activity Feeds Architecture
January, 2011

Friday, January 14, 2011


To Be Covered:
•Data model
•Where feeds come from
•How feeds are displayed
•Optimizations



Fundamental Entities
Connections Activities


There are two fundamental building blocks for feeds: connections and activities.
Activities form a log of what some entity on the site has done, or had done to it.
Connections express relationships between entities.
I will explain the data model for connections ﬁrst.


Connections

Favorites
Circles

Connections

Etc.

Orders


Connections are a superset of Circles, Favorites, Orders, and other relationships between
entities on the site.


Connections
A B

C

F

E D

G I

H

J


Connections are implemented as a directed graph.
Currently, the nodes can be people or shops. (In principle they can be other objects.)


Connections
A B

C

F
connection_edges_forward

connection_edges_reverse E D

G I

H

J


The edges of the graph are stored in two tables.
For any node, connection_edges_forward lists outgoing edges and connection_edges_reverse
lists the incoming edges.
In other words, we store each edge twice.


Connections
aBA
aCB

A B

aBC

aFA aEA C
aEB
aDB
aCD

aEF
F

aFE
E D

aEI
aDI

aIG
aHE

G aJG I

aHG aGJ
aIJ

H aHJ

J


We also assign each edge a weight, known as affinity.


Connections
On H’s shard

connection_edges_forward

from to afﬁnity
H E 0.3
H G 0.7
connection_edges_reverse

from to afﬁnity
J H 0.75


Here we see the data for Anda’s connections on her shard.
She has two entries in the forward connections table for the people in her circle.
She has one entry in the reverse connections so that she can see everyone following her.


Activities


Activities are the other database entity important to activity feeds.


activity := (subject, verb, object)


As you can see in Rob’s magnetic poetry diagram, activities are a description of an event on
Etsy boiled down to a subject (“who did it”), a verb (“what they did”), and an object (“what
they did it to”).



(Steve, connected, Kyle)

(Kyle, favorited, brief jerky)


Here are some examples of activities.
The ﬁrst one describes Steve adding Kyle to his circle.
The second one describes Kyle favoriting an item.
In each of these cases note that there are probably several parties interested in these events
[examples]. The problem (the main one we’re trying to solve with activity feeds) is how to
notify all of them about it. In order to achieve that goal, as usual we copy the data all over the
place.



activity := [owner,(subject, verb, object)]


So what we do is duplicate the S,V,O combinations with different owners.
Steve will have his record that he connected to Kyle, and Kyle will be given his own record
that Steve connected to him.



[Steve, (Steve, connected, Kyle)] Steve's shard


[Kyle, (Steve, connected, Kyle)] Kyle's shard


This is what that looks like.



[Kyle, (Kyle, favorited, brief jerky)] Kyle's shard


[MixedSpecies, (Kyle, favorited, brief jerky)]
Mixedspecies' shard
[brief jerky, (Kyle, favorited, brief jerky)]


In more complicated examples there could be more than two owners.
You could envision people being interested in Kyle, people being interested in MixedSpecies,
or people being interested in brief jerky.
In cases where there are this many writes, we will generally perform them with Gearman.
Again, in order for interested parties to ﬁnd the activities, we copy the activities all over the
place.


Building a Feed

‫ּט_ּט‬

Magic,
cheating

Magic,
Newsfeed
cheating


Now that we know about connections and activities, we can talk about how activities are
turned into Newsfeeds and how those wind up being displayed to end users.


Building a Feed

‫ּט_ּט‬
Display

Magic,
cheating

Magic,
Newsfeed
cheating

Aggregation


Getting to the end result (the activity feed page) has two distinct phases: aggregation and
display.


Aggregation

Shard 1 (Steve, favorited, foo)

(theblackapple, listed, widget)
(Wil, bought, mittens)
Shard 2

Newsfeed

(Wil, bought, mittens)
Shard 3

Shard 4 (Kyle, favorited, brief jerky)


I am going to talk about aggregation ﬁrst.
Aggregation turns activities (in the database) into a Newsfeed (in memcache).
Aggregation typically occurs offline, with Gearman.


Aggregation, Step 1: Choosing Connections

Potentially way too many


We already allow people to have more connections than would make sense on a single feed,
or could be practically aggregated all at once.
The ﬁrst step in aggregation is to turn the list of people you are connected to into the list of
people we’re actually going to go seek out activities for.



“In Theory”

1

0.75
Affinity
0.5

0.25

0
Connection

$choose_connection = mt_rand() < $affinity;


In theory, the way we would do this is rank the connections by affinity and then treat the
affinity as the probability that we’ll pick it.



“In Theory”

1

0.75
Affinity
0.5

0.25

0
Connection

$choose_connection = mt_rand() < $affinity;


So then we’d be more likely to pick the close connections, but leaving the possibility that we
will pick the distant ones.



“In Practice”

1

0.75
Affinity
0.5

0.25

0
Connection


In practice we don’t really handle affinity yet.



“Even More In Practice”

1

0.75
Affinity
0.5

0.25

0

Connection


And most people don’t currently have enough connections for this to matter at all. (Mean
connections is around a dozen.)


Aggregation, Step 2: Making Activity Sets

score activity activity activity activity

0.0 0.0 0.0 0.0
flags
0x0 0x0 0x0 0x0

activity activity

0.0 0.0
0x0 0x0

activity activity activity

0.0 0.0 0.0
0x0 0x0 0x0

activity activity activity activity

0.0 0.0 0.0 0.0
0x0 0x0 0x0 0x0


Once the connections are chosen, we then select historical activity for them and convert them
into in-memory structures called activity sets.
These are just the activities grouped by connection, with a score and flags field for each.
The next few phases of aggregation operate on these.


Aggregation, Step 3: Classification


0.0 0.0 0.0 0.0
0x11 0x3 0x20c1 0x10004

activity activity

0.0 0.0
0x20c1 0x4


0.0 0.0 0.0
0x1001 0x2003 0x11


0.0 0.0 0.0 0.0
0x11 0x401 0x5 0x10004


The next thing that happens is that we iterate through all of the activities in all of the sets and
classify them.


Aggregation, Step 3: Classification


0.0 0.0 0.0 0.0
0x11 0x3 0x20c1 0x10004 about_owner_shop | user_created_treasury

Rob created a treasury featuring
the feed owner's shop.

activity activity

0.0 0.0
0x20c1 0x4


0.0 0.0 0.0
0x1001 0x2003 0x11


0.0 0.0 0.0 0.0
0x11 0x401 0x5 0x10004


The flags are a bit field.
They are all from the point of view of the feed owner.
So the same activities on another person’s feed would be assigned different flags.


Aggregation, Step 4: Scoring


0.8 0.77 0.9 0.1
0x11 0x3 0x20c1 0x10004

activity activity

0.6 0.47
0x20c1 0x4


0.8 0.55 0.8
0x1001 0x2003 0x11


0.3 0.25 0.74 0.9
0x11 0x401 0x5 0x10004


Next we ﬁll in the score ﬁelds.
At this point the score is just a simple time decay function (older activities always score lower
than new ones).


Aggregation, Step 5: Pruning
[Rob, (Rob, connected, Jared)]


0.8 0.77 0.9 0.1
0x11 0x3 0x20c1 0x10004

activity activity

0.6 0.47
0x20c1 0x4

activity activity activity [Jared, (Rob, connected, Jared)]

0.8 0.55 0.8
0x1001 0x2003 0x11


0.3 0.25 0.74 0.9
0x11 0x401 0x5 0x10004


As we noted before it’s possible to wind up seeing the same event as two or more activities.
The next stage of aggregation detects these situations.


Aggregation, Step 5: Pruning
[Rob, (Rob, connected, Jared)]


0.8 0.77 0.9 0.1
0x11 0x3 0x20c1 0x10004

activity activity

0.6 0.47
0x20c1 0x4

activity activity activity [Jared, (Rob, connected, Jared)]

0.8 0.55 0.8
0x1001 0x2003 0x11


0.3 0.25 0.74 0.9
0x11 0x401 0x5 0x10004


We iterate through the activity sets and remove the duplicates.
Right now we can just cross off the second instance of the SVO pair; once we have comments
this will be more complicated.


Aggregation, Step 6: Sort & Merge


0.8 0.77 0.9 0.1
0x11 0x3 0x20c1 0x10004

activity activity

0.6 0.47
0x20c1 0x4
activity activity activity activity activity activity activity activity activity activity activity activity

0.9 0.9 0.8 0.8 0.77 0.74 0.6 0.55 0.47 0.3 0.25 0.1
0x20c1 0x10004 0x1001 0x11 0x3 0x5 0x20c1 0x2003 0x4 0x11 0x401 0x10004
activity activity

0.8 0.55
0x1001 0x2003


0.3 0.25 0.74 0.9
0x11 0x401 0x5 0x10004


Once everything is scored, classiﬁed, and de-duped we can ﬂatten the whole thing and sort it
by score.


Aggregation, Step 6: Sort & Merge

activity activity activity activity activity activity activity activity activity activity activity activity

0.9 0.9 0.8 0.8 0.77 0.74 0.6 0.55 0.47 0.3 0.25 0.1
0x20c1 0x10004 0x1001 0x11 0x3 0x5 0x20c1 0x2003 0x4 0x11 0x401 0x10004

activity activity activity activity activity activity activity activity activity activity activity activity activity activity activity activity activity activity activity

0.9 0.9 0.8 0.8 0.77 0.74 0.716 0.71 0.6 0.6 0.55 0.47 0.3 0.3 0.291 0.25 0.1 0.097 0.02
0x20c1 0x10004 0x1001 0x11 0x3 0x5 0x4c01 0x2c01 0x20c1 0xc001 0x2003 0x4 0x11 0x4001 0x4001 0x401 0x10004 0x4 0x1001

Newsfeed


Then we take the ﬁnal set of activities and merge it on to the owner’s existing newsfeed.
(Or we create a new newsfeed if they don’t have one.)


Aggregation: Cleaning Up

Just ﬁne Too many

activity activity activity activity activity activity activity activity activity activity activity activity activity activity activity activity activity activity activity

0.9 0.9 0.8 0.8 0.77 0.74 0.716 0.71 0.6 0.6 0.55 0.47 0.3 0.3 0.291 0.25 0.1 0.097 0.02
0x20c1 0x10004 0x1001 0x11 0x3 0x5 0x4c01 0x2c01 0x20c1 0xc001 0x2003 0x4 0x11 0x4001 0x4001 0x401 0x10004 0x4 0x1001

Newsfeed


We trim off the end of the newsfeed, so that they don’t become arbitrarily large.


Aggregation: Cleaning Up

activity activity activity activity activity activity activity activity activity activity activity activity activity activity

0.9 0.9 0.8 0.8 0.77 0.74 0.716 0.71 0.6 0.6 0.55 0.47 0.3 0.3
0x20c1 0x10004 0x1001 0x11 0x3 0x5 0x4c01 0x2c01 0x20c1 0xc001 0x2003 0x4 0x11 0x4001

Newsfeed

memcached


And then ﬁnally we stuff the feed into memcached.


Aggregation


Currently we peak at doing this about 25 times per second.


Aggregation


We do it for a lot of different reasons:
- The feed owner has done something, or logged in.
- On a schedule, with cron.
- We also aggregate for your connections when you do something (purple). This is a hack and
won’t scale.


Display

memcached


Next I’m going to talk about how we get from the memcached newsfeed to the ﬁnal product.


Display: done naively, sucks


If we just showed the activities in the order that they’re in on the newsfeed, it would look like
this.


Display: Enter Rollups


To solve this problem we combine similar stories into rollups.


Display: Computing Rollups


Here’s an attempt at depicting how rollups are created.
The feed is divided up into sections, so that you don’t wind up seeing all of the reds, greens,
etc. on the entire feed in just a few very large rollups.
Then the similar stories are grouped together within the sections.


Display: Filling in Stories

activity Story
Story
0.9
html
(Feed-owner-
StoryHydrator (Global details) StoryTeller Smarty
0x10004 speciﬁc details)

memcached


Once that’s done, we can go through the rest of the display pipeline for the root story in each
rollup.
There are multiple layers of caching here. Things that are global (like the shop associated
with a favorited listing) are cached separately from things that are unique to the person
looking at the feed (like the exact way the story is phrased).


Making it Fast
Response Time (ms)

1200

900

Boom
600

300

0
Homepage Shop Listing Search Activity


Finally I’m going to go through a few ways that we’ve sped up activity, to the point where it’s
one of the faster pages on the site (despite being pretty complicated).


Hack #1: Cache Warming
!_!


Magic,
cheating

Magic,
Newsfeed
cheating


The ﬁrst thing we do to speed things up is run almost the entire pipeline proactively using
gearman.
So after aggregation we trigger a display run, even though nobody is there to look at the
html.
The end result is that almost every pageview is against a hot cache.


Hack #2: TTL Caching

May be his avatar
from 5 minutes
ago.

Big f’ing deal.


The second thing we do is add bits of TTL caching where few people will notice.
Straightforward but not done in many places on the site.
Note that his avatar here is tied to the story. If he generates new activity he’ll see his new
avatar.


Hack #3: Judicious Associations

getFinder(“UserProfile”)->find
(...)

not
getFinder(“User”)->find(...)-
>Profile


We also profiled the pages and meticulously simplified ORM usage.
Again this sounds obvious but it’s really easy to lose track of what you’re doing as you hand
the user off to the template. Lots of ORM calls were originally actually being made by the
template.


Hack #4: Lazy Below the Fold

We don’t load much at the outset.
You get more as you scroll down
(ﬁnite scrolling).


The End


Etsy Activity Feeds Architecture

More Related Content

What's hot (20)

Viewers also liked (11)

Similar to Etsy Activity Feeds Architecture (20)

Recently uploaded (20)

Etsy Activity Feeds Architecture