1
00:00:00,040 --> 00:00:03,600
Welcome to the M365 FM podcast,
where we explore the people,

2
00:00:03,600 --> 00:00:06,519
ideas and technologies shaping
the future of work.

3
00:00:06,519 --> 00:00:08,439
What a semantic model actually
is.

4
00:00:08,439 --> 00:00:09,679
A semantic model is one of those

5
00:00:09,679 --> 00:00:11,119
terms that makes beginners close

6
00:00:11,119 --> 00:00:11,519
the tab.

7
00:00:11,519 --> 00:00:13,839
It sounds like something only a
database administrator would

8
00:00:13,839 --> 00:00:16,280
need, but it's actually the
layer that decides what every

9
00:00:16,280 --> 00:00:17,920
number in your reports means.

10
00:00:17,920 --> 00:00:21,179
Microsoft gave it a new name
because the old word dataset

11
00:00:21,179 --> 00:00:22,600
never explained anything.

12
00:00:22,600 --> 00:00:25,800
Fabric made that layer something
you build and reuse, and it

13
00:00:25,800 --> 00:00:27,359
changed how the data gets in.

14
00:00:27,359 --> 00:00:28,480
By the end of this, you'll know

15
00:00:28,480 --> 00:00:30,039
what it is, how it works, where

16
00:00:30,039 --> 00:00:31,379
it breaks, and where it sits

17
00:00:31,379 --> 00:00:32,759
next to teams, Excel and

18
00:00:32,759 --> 00:00:33,520
copilot.

19
00:00:33,520 --> 00:00:37,079
Welcome to the M365 FM podcast
where we explore the people,

20
00:00:37,079 --> 00:00:39,799
ideas, and technologies shaping
the future of work.

21
00:00:39,799 --> 00:00:41,759
So let's start with the plain
version.

22
00:00:41,759 --> 00:00:43,679
Because the name does most of
the damage.

23
00:00:43,679 --> 00:00:46,399
It's a description of one part
of your business written in the

24
00:00:46,399 --> 00:00:47,920
words people already use.

25
00:00:47,920 --> 00:00:49,280
Microsoft calls it a logical

26
00:00:49,280 --> 00:00:50,600
description of an analytical

27
00:00:50,600 --> 00:00:52,439
domain with metrics and business

28
00:00:52,439 --> 00:00:53,280
friendly names.

29
00:00:53,280 --> 00:00:54,960
That phrase is hiding something
simple.

30
00:00:54,960 --> 00:00:57,880
Take one number your team argues
about every month, like revenue.

31
00:00:57,880 --> 00:00:59,280
Now count how many ways for

32
00:00:59,280 --> 00:01:00,560
different people would calculate

33
00:01:00,560 --> 00:01:00,960
it.

34
00:01:00,960 --> 00:01:03,359
The semantic model is where you
settle that argument once by

35
00:01:03,359 --> 00:01:05,920
defining what revenue means,
which rows count, and how it

36
00:01:05,920 --> 00:01:07,200
connects to everything else.

37
00:01:07,200 --> 00:01:09,519
That sounds abstract, so let's
make it concrete.

38
00:01:09,519 --> 00:01:12,120
When the definition lives in one
place, you change it once and

39
00:01:12,120 --> 00:01:13,359
every report updates.

40
00:01:13,359 --> 00:01:14,480
That beats hunting through a

41
00:01:14,480 --> 00:01:15,579
dozen spreadsheets that each

42
00:01:15,579 --> 00:01:16,719
invented their own version of

43
00:01:16,719 --> 00:01:17,439
the truth.

44
00:01:17,439 --> 00:01:18,680
If you've used power BI, you

45
00:01:18,680 --> 00:01:20,120
already know this object under a

46
00:01:20,120 --> 00:01:21,000
different name.

47
00:01:21,000 --> 00:01:22,680
Microsoft used to call it a data
set.

48
00:01:22,680 --> 00:01:24,840
The name changed because data
set was too generic.

49
00:01:24,840 --> 00:01:26,319
The new one signals a reusable

50
00:01:26,319 --> 00:01:27,959
asset, not a file bolted onto a

51
00:01:27,959 --> 00:01:28,879
single report.

52
00:01:28,879 --> 00:01:32,200
Plenty of older tutorials still
say data set, so you'll see both

53
00:01:32,200 --> 00:01:34,760
words in the wild, and they
point at the same thing.

54
00:01:34,760 --> 00:01:37,319
The cleanest way to hold it in
your head is this.

55
00:01:37,319 --> 00:01:40,159
The semantic model carries your
business definitions, and it

56
00:01:40,159 --> 00:01:42,640
sits between your raw data and
everything built on top of it.

57
00:01:42,640 --> 00:01:44,819
But a model needs data
underneath it, and that's where

58
00:01:44,819 --> 00:01:46,159
fabric comes in.

59
00:01:46,159 --> 00:01:49,420
How power BI and fabric fit
together fabric is Microsoft's

60
00:01:49,420 --> 00:01:51,359
all in one analytics platform.

61
00:01:51,359 --> 00:01:54,359
The easiest way to picture it is
one big data lake with a set of

62
00:01:54,359 --> 00:01:55,560
tools sitting on top.

63
00:01:55,560 --> 00:01:57,480
That lake has a name, one lake

64
00:01:57,480 --> 00:01:59,120
and every fabric tenant gets one

65
00:01:59,120 --> 00:01:59,640
automatically.

66
00:01:59,640 --> 00:02:00,840
Whether or not anyone asked for

67
00:02:00,840 --> 00:02:02,319
it, you can't delete it, and you

68
00:02:02,319 --> 00:02:03,760
can't create a second one that

69
00:02:03,760 --> 00:02:05,000
tells you a lot about the intent

70
00:02:05,000 --> 00:02:05,519
behind it.

71
00:02:05,519 --> 00:02:08,479
Because Microsoft wants your
organization's data in one place

72
00:02:08,479 --> 00:02:11,000
instead of scattered across a
dozen storage accounts.

73
00:02:11,000 --> 00:02:13,879
Fabric organizes all of this
into a simple hierarchy.

74
00:02:13,879 --> 00:02:16,840
It's worth 30s because the same
words show up everywhere.

75
00:02:16,840 --> 00:02:19,919
The tenant sits at the top and
maps to your whole organization.

76
00:02:19,919 --> 00:02:22,599
Under that come workspaces,
which act like shared folders

77
00:02:22,599 --> 00:02:24,080
for a team or a project.

78
00:02:24,080 --> 00:02:25,659
Inside a workspace, you find

79
00:02:25,659 --> 00:02:27,400
items and a lake house is one of

80
00:02:27,400 --> 00:02:28,879
them, a place where files and

81
00:02:28,879 --> 00:02:30,360
tables live together like a

82
00:02:30,360 --> 00:02:31,800
database sitting on top of a

83
00:02:31,800 --> 00:02:33,240
data lake.

84
00:02:33,240 --> 00:02:35,800
One lake is built on Azure Data
Lake storage.

85
00:02:35,800 --> 00:02:37,240
It keeps tables in open formats

86
00:02:37,240 --> 00:02:38,319
called delta, parquet and

87
00:02:38,319 --> 00:02:39,080
iceberg.

88
00:02:39,080 --> 00:02:42,240
Those two words matter more than
they sound, because open formats

89
00:02:42,240 --> 00:02:44,080
mean any tool can read the data.

90
00:02:44,080 --> 00:02:46,000
You're never locked into one
product forever.

91
00:02:46,000 --> 00:02:47,840
Power BI lives inside fabric as

92
00:02:47,840 --> 00:02:49,199
the reporting tool, and the

93
00:02:49,199 --> 00:02:50,280
semantic model sits in the

94
00:02:50,280 --> 00:02:50,800
middle.

95
00:02:50,800 --> 00:02:52,000
Data lands in the lake.

96
00:02:52,000 --> 00:02:54,879
The model describes what the
data means and reports read from

97
00:02:54,879 --> 00:02:56,280
the model before fabric.

98
00:02:56,280 --> 00:02:58,120
That model lived inside power BI

99
00:02:58,120 --> 00:02:59,479
attached to whichever report

100
00:02:59,479 --> 00:03:01,039
created it fabric, pulled it

101
00:03:01,039 --> 00:03:02,360
out, and made it a standalone

102
00:03:02,360 --> 00:03:03,719
item with its own name and its

103
00:03:03,719 --> 00:03:04,560
own life.

104
00:03:04,560 --> 00:03:05,879
One model can now serve many

105
00:03:05,879 --> 00:03:07,360
reports instead of hiding inside

106
00:03:07,360 --> 00:03:07,800
one.

107
00:03:07,800 --> 00:03:10,719
That's the setup, and it's a lot
of moving parts to hold at once.

108
00:03:10,719 --> 00:03:12,319
So let's open the model itself

109
00:03:12,319 --> 00:03:14,280
and look at what's inside the

110
00:03:14,280 --> 00:03:16,159
building blocks inside a model,

111
00:03:16,159 --> 00:03:17,680
open up a semantic model and you

112
00:03:17,680 --> 00:03:19,039
find three things which are

113
00:03:19,039 --> 00:03:20,599
tables, relationships, and

114
00:03:20,599 --> 00:03:21,319
measures.

115
00:03:21,319 --> 00:03:24,080
The tables usually follow a
pattern called a star schema.

116
00:03:24,080 --> 00:03:26,479
That name sounds mathematical
when it's really just a way of

117
00:03:26,479 --> 00:03:28,879
separating what happened from
how you want to look at it.

118
00:03:28,879 --> 00:03:32,000
The tables that record events
are called facts, and each row

119
00:03:32,000 --> 00:03:34,879
is something that actually
occurred like a sale, an order,

120
00:03:34,879 --> 00:03:36,240
or a support ticket.

121
00:03:36,240 --> 00:03:37,439
The tables that let you slice

122
00:03:37,439 --> 00:03:38,400
those events are called

123
00:03:38,400 --> 00:03:40,120
dimensions, and they hold the

124
00:03:40,120 --> 00:03:41,479
lenses you care about like

125
00:03:41,479 --> 00:03:43,240
product, region, customer, and

126
00:03:43,240 --> 00:03:43,840
date.

127
00:03:43,840 --> 00:03:45,479
A relationship connects the two.

128
00:03:45,479 --> 00:03:48,319
Each fact row points back to the
dimensions it belongs to.

129
00:03:48,319 --> 00:03:51,639
So a single sale can be tied to
a product, a store, and a day.

130
00:03:51,639 --> 00:03:54,409
Once those links exist, you can
ask questions that span them.

131
00:03:54,409 --> 00:03:57,129
You could check how many units
of a specific product sold in a

132
00:03:57,129 --> 00:03:58,810
specific region last quarter.

133
00:03:58,810 --> 00:04:00,050
Then there are the measures, and

134
00:04:00,050 --> 00:04:01,289
this is where the model earns

135
00:04:01,289 --> 00:04:02,009
its keep.

136
00:04:02,009 --> 00:04:03,090
A measure is a calculation

137
00:04:03,090 --> 00:04:04,129
written in a language called

138
00:04:04,129 --> 00:04:05,770
Dax, and it turns rows into

139
00:04:05,770 --> 00:04:06,210
numbers.

140
00:04:06,210 --> 00:04:10,449
Total revenue units sold and
profit margin are all measures.

141
00:04:10,449 --> 00:04:12,770
Take a simple one to get total
revenue.

142
00:04:12,770 --> 00:04:15,250
A measure adds up a revenue
column across all the rows.

143
00:04:15,250 --> 00:04:17,850
A report is showing it
recalculates itself the moment

144
00:04:17,850 --> 00:04:19,129
someone clicks a filter.

145
00:04:19,129 --> 00:04:21,370
That's the part people
underestimate because a measure

146
00:04:21,370 --> 00:04:22,449
behaves like a rule.

147
00:04:22,449 --> 00:04:25,250
It responds to whatever slice of
data is on screen instead of

148
00:04:25,250 --> 00:04:26,810
holding a frozen number.

149
00:04:26,810 --> 00:04:28,209
There's also a middle layer

150
00:04:28,209 --> 00:04:29,170
worth knowing, called a

151
00:04:29,170 --> 00:04:30,490
calculated column.

152
00:04:30,490 --> 00:04:33,290
That's a value worked out row by
row and stored in the table so

153
00:04:33,290 --> 00:04:36,329
it behaves more like a normal
column than a live calculation.

154
00:04:36,329 --> 00:04:38,649
Keeping columns and measures
straight in your head saves a

155
00:04:38,649 --> 00:04:39,930
lot of confusion later.

156
00:04:39,930 --> 00:04:43,149
The quiet superpower here is
naming a column called rev

157
00:04:43,149 --> 00:04:44,610
underscore edge underscore.

158
00:04:44,610 --> 00:04:47,810
LC can become net revenue after
currency adjustment.

159
00:04:47,810 --> 00:04:48,850
That single rename is the

160
00:04:48,850 --> 00:04:49,970
difference between a report

161
00:04:49,970 --> 00:04:51,589
people trust and a report They

162
00:04:51,589 --> 00:04:53,129
squint at tables and columns,

163
00:04:53,129 --> 00:04:54,449
get human names, descriptions

164
00:04:54,449 --> 00:04:55,889
get filled in, and the data

165
00:04:55,889 --> 00:04:57,129
starts speaking the language of

166
00:04:57,129 --> 00:04:57,810
the business.

167
00:04:57,810 --> 00:05:00,649
Hold on to that idea, because
the shape of the model is only

168
00:05:00,649 --> 00:05:01,610
half the story.

169
00:05:01,610 --> 00:05:04,009
The other half is how the data
actually gets in, and that

170
00:05:04,009 --> 00:05:07,129
choice changes everything about
how fast the model feels.

171
00:05:07,129 --> 00:05:09,209
The three ways data gets into a
model.

172
00:05:09,209 --> 00:05:10,230
There are three main ways to

173
00:05:10,230 --> 00:05:11,930
feed a semantic model, and they

174
00:05:11,930 --> 00:05:13,649
trade freshness against speed in

175
00:05:13,649 --> 00:05:14,209
different ways.

176
00:05:14,209 --> 00:05:15,730
The first is import, which

177
00:05:15,730 --> 00:05:17,129
copies the data into the model's

178
00:05:17,129 --> 00:05:17,449
memory.

179
00:05:17,449 --> 00:05:20,649
When it refreshes, queries fly
after that because everything is

180
00:05:20,649 --> 00:05:21,810
already loaded in memory.

181
00:05:21,810 --> 00:05:23,009
The catch is that the data is

182
00:05:23,009 --> 00:05:24,370
only as current as the last

183
00:05:24,370 --> 00:05:25,129
refresh.

184
00:05:25,129 --> 00:05:27,610
If your refresh runs at six in
the morning, your numbers are a

185
00:05:27,610 --> 00:05:29,089
snapshot of six in the morning.

186
00:05:29,089 --> 00:05:31,250
The second is direct query,
which does the opposite.

187
00:05:31,250 --> 00:05:33,610
It leaves the data where it
lives and queries the source.

188
00:05:33,610 --> 00:05:35,209
Every time someone opens a
report.

189
00:05:35,209 --> 00:05:37,490
You always see the latest
numbers, but the source database

190
00:05:37,490 --> 00:05:40,250
usually wasn't built to answer
thousands of report queries.

191
00:05:40,250 --> 00:05:42,370
Everything slows down when
traffic picks up.

192
00:05:42,370 --> 00:05:43,649
The third one is the interesting

193
00:05:43,649 --> 00:05:45,149
one, and it only exists in

194
00:05:45,149 --> 00:05:45,649
fabric.

195
00:05:45,649 --> 00:05:47,009
It's called direct lake.

196
00:05:47,009 --> 00:05:48,170
Instead of copying data or

197
00:05:48,170 --> 00:05:49,689
querying a database, it reads

198
00:05:49,689 --> 00:05:51,089
the delta parquet files in one

199
00:05:51,089 --> 00:05:51,769
lake directly.

200
00:05:51,769 --> 00:05:52,850
There's no copy step, no

201
00:05:52,850 --> 00:05:53,930
overnight refresh, and no

202
00:05:53,930 --> 00:05:54,930
separate database sitting

203
00:05:54,930 --> 00:05:55,410
between them.

204
00:05:55,410 --> 00:05:58,449
You get the speed of import with
the freshness of direct query,

205
00:05:58,449 --> 00:06:01,129
which is the combination that
used to be impossible.

206
00:06:01,129 --> 00:06:03,850
Choosing between them comes down
to a few practical things.

207
00:06:03,850 --> 00:06:07,009
Import is the safe default for
small and medium models because

208
00:06:07,009 --> 00:06:09,930
it's fast, flexible, and works
on any license.

209
00:06:09,930 --> 00:06:12,769
Direct query makes sense when
the data has to be live to the

210
00:06:12,769 --> 00:06:15,889
second, or when it lives outside
fabric in something like an

211
00:06:15,889 --> 00:06:19,329
Azure SQL database or snowflake
Direct Lake is built for the

212
00:06:19,329 --> 00:06:22,290
opposite end of the scale, where
the tables are huge and the data

213
00:06:22,290 --> 00:06:24,569
is already sitting in one lake
at that size.

214
00:06:24,569 --> 00:06:27,529
Copying all of it would be slow,
expensive, or simply impossible.

215
00:06:27,529 --> 00:06:28,850
You can also mix them in.

216
00:06:28,850 --> 00:06:31,810
A single model, can hold a giant
fact table in Direct Lake and a

217
00:06:31,810 --> 00:06:33,250
small lookup table in import.

218
00:06:33,250 --> 00:06:35,129
That helps when the small table
needs a bit of cleaning.

219
00:06:35,129 --> 00:06:36,889
That only Power Query can do.

220
00:06:36,889 --> 00:06:39,209
Fabric calls that a composite
model, and it's a practical

221
00:06:39,209 --> 00:06:41,649
answer to a real problem,
because most companies don't get

222
00:06:41,649 --> 00:06:43,970
to pick one clean approach for
everything they own.

223
00:06:43,970 --> 00:06:46,930
That sounds almost too good, and
it's fair to be skeptical.

224
00:06:46,930 --> 00:06:49,410
The mechanism behind it comes
down to two words you'll hear

225
00:06:49,410 --> 00:06:51,290
from anyone who works in fabric.

226
00:06:51,290 --> 00:06:54,329
Together, they explain why a
model can stay fast even when

227
00:06:54,329 --> 00:06:56,569
the data is enormous.

228
00:06:56,569 --> 00:06:58,410
How direct Lake actually works.

229
00:06:58,410 --> 00:06:59,930
The first word is transcoding,

230
00:06:59,930 --> 00:07:01,370
and it describes how the model

231
00:07:01,370 --> 00:07:02,370
loads data.

232
00:07:02,370 --> 00:07:04,689
When a report asks a question,
the model works out which

233
00:07:04,689 --> 00:07:05,529
columns it needs.

234
00:07:05,529 --> 00:07:08,370
Then it loads only those columns
from one lake into memory.

235
00:07:08,370 --> 00:07:11,209
If your report shows revenue by
region, it loads the revenue

236
00:07:11,209 --> 00:07:13,410
column and the region column,
and it leaves the other one

237
00:07:13,410 --> 00:07:14,490
hundred columns alone.

238
00:07:14,490 --> 00:07:17,089
The next time someone asks a
similar question, those columns

239
00:07:17,089 --> 00:07:19,930
are already in memory and the
answer comes back instantly.

240
00:07:19,930 --> 00:07:21,529
The second word is framing, and

241
00:07:21,529 --> 00:07:22,649
it's fabric's version of a

242
00:07:22,649 --> 00:07:23,290
refresh.

243
00:07:23,290 --> 00:07:26,769
Instead of copying data, framing
reads the metadata of the delta

244
00:07:26,769 --> 00:07:30,329
tables, notices which files
changed and updates the model's

245
00:07:30,329 --> 00:07:31,930
pointers to the latest ones.

246
00:07:31,930 --> 00:07:33,810
That usually takes seconds
because it's moving.

247
00:07:33,810 --> 00:07:36,490
Bookkeeping rather than data
framing, marks the moment your

248
00:07:36,490 --> 00:07:39,569
model trusts, so the data you
see is as of the last successful

249
00:07:39,569 --> 00:07:41,569
framing that detail matters more
than it sounds.

250
00:07:41,569 --> 00:07:44,329
There are two versions of Direct
Lake, and the difference matters

251
00:07:44,329 --> 00:07:45,569
when something goes wrong.

252
00:07:45,569 --> 00:07:47,009
Direct leg on one leg, reads the

253
00:07:47,009 --> 00:07:48,490
leg files directly and has no

254
00:07:48,490 --> 00:07:49,089
fallback.

255
00:07:49,089 --> 00:07:51,490
If a query can't run, it fails
and someone fixes it.

256
00:07:51,490 --> 00:07:52,730
Direct leg on sequel goes

257
00:07:52,730 --> 00:07:54,290
through a SQL endpoint instead,

258
00:07:54,290 --> 00:07:55,410
and when it hits something it

259
00:07:55,410 --> 00:07:56,649
can't handle, it falls back to

260
00:07:56,649 --> 00:07:57,410
Direct query.

261
00:07:57,410 --> 00:08:00,129
That fallback keeps reports
alive, but it also means a

262
00:08:00,129 --> 00:08:02,490
report can slow down without
anyone changing a thing.

263
00:08:02,490 --> 00:08:05,250
There's a setting called keep
your direct leg data up to date,

264
00:08:05,250 --> 00:08:06,529
and it's on by default.

265
00:08:06,529 --> 00:08:07,930
It watches the lake and reframes

266
00:08:07,930 --> 00:08:08,970
automatically when something

267
00:08:08,970 --> 00:08:09,990
changes, so you don't have to

268
00:08:09,990 --> 00:08:11,350
babysit a refresh schedule

269
00:08:11,350 --> 00:08:12,990
because only the columns a query

270
00:08:12,990 --> 00:08:14,449
needs are pulled into memory,

271
00:08:14,449 --> 00:08:16,689
the total data can be far larger

272
00:08:16,689 --> 00:08:18,009
than the memory your capacity

273
00:08:18,009 --> 00:08:18,649
provides.

274
00:08:18,649 --> 00:08:21,769
Freshness has one more wrinkle
worth knowing because framing

275
00:08:21,769 --> 00:08:24,490
locks in a moment, the model
won't show you a data load

276
00:08:24,490 --> 00:08:25,689
that's only half finished.

277
00:08:25,689 --> 00:08:27,850
That's usually what you want,
since a stable number beats a

278
00:08:27,850 --> 00:08:29,970
number that keeps moving while
someone is reading it.

279
00:08:29,970 --> 00:08:32,690
Put those pieces together and
the pitch starts to make sense.

280
00:08:32,690 --> 00:08:34,090
You point the model at the lake.

281
00:08:34,090 --> 00:08:35,570
It grabs only what each question

282
00:08:35,570 --> 00:08:37,049
needs and it stays current

283
00:08:37,049 --> 00:08:38,370
without copying the world every

284
00:08:38,370 --> 00:08:39,289
night.

285
00:08:39,289 --> 00:08:40,769
What this looks like in
practice.

286
00:08:40,769 --> 00:08:42,210
Picture a concrete version, the

287
00:08:42,210 --> 00:08:43,330
kind of thing you do on your

288
00:08:43,330 --> 00:08:44,009
first day.

289
00:08:44,009 --> 00:08:45,289
Imagine a sales team with order

290
00:08:45,289 --> 00:08:46,370
data landing in a fabric

291
00:08:46,370 --> 00:08:47,049
lakehouse.

292
00:08:47,049 --> 00:08:50,090
Someone opens that lakehouse,
chooses new power BI semantic

293
00:08:50,090 --> 00:08:53,210
model, gives it a name, and
picks the tables to include.

294
00:08:53,210 --> 00:08:55,850
From there, they add the
relationships between orders,

295
00:08:55,850 --> 00:08:57,169
products, and dates.

296
00:08:57,169 --> 00:08:59,450
Then they write a handful of
measures like total revenue and

297
00:08:59,450 --> 00:09:00,690
average order value.

298
00:09:00,690 --> 00:09:01,490
That's the whole build.

299
00:09:01,490 --> 00:09:04,090
And once it exists, reports are
the easy part.

300
00:09:04,090 --> 00:09:06,889
You can create one right in the
browser or open Power BI

301
00:09:06,889 --> 00:09:09,730
desktop, connect to the model
and build there instead.

302
00:09:09,730 --> 00:09:12,549
The important shift is that the
model is the thing people share,

303
00:09:12,549 --> 00:09:13,730
not the report.

304
00:09:13,730 --> 00:09:16,409
One model can feed a finance
dashboard, an operations

305
00:09:16,409 --> 00:09:18,250
scorecard, and a sales report.

306
00:09:18,250 --> 00:09:19,929
All of them pull the same
definitions.

307
00:09:19,929 --> 00:09:23,330
So a number means the same thing
everywhere it appears publishing

308
00:09:23,330 --> 00:09:24,450
follows the same logic.

309
00:09:24,450 --> 00:09:25,769
You save the model to a

310
00:09:25,769 --> 00:09:27,289
workspace, set permissions for

311
00:09:27,289 --> 00:09:28,570
the people who need it, and

312
00:09:28,570 --> 00:09:30,049
anyone with access can build on

313
00:09:30,049 --> 00:09:30,450
it.

314
00:09:30,450 --> 00:09:32,850
The model becomes the shared
source and the reports become

315
00:09:32,850 --> 00:09:34,250
the parts you swap out.

316
00:09:34,250 --> 00:09:35,210
That's the order you want.

317
00:09:35,210 --> 00:09:37,769
As a team grows and the
questions start changing.

318
00:09:37,769 --> 00:09:39,250
There's one small warning here.

319
00:09:39,250 --> 00:09:41,629
It's tempting to skip the
descriptions and leave the

320
00:09:41,629 --> 00:09:43,570
cryptic column names in place.

321
00:09:43,570 --> 00:09:46,129
That shortcut always comes back
to bite you, because the next

322
00:09:46,129 --> 00:09:49,009
person to open the model has no
idea what anything means.

323
00:09:49,009 --> 00:09:50,250
This is also where the beginner

324
00:09:50,250 --> 00:09:51,409
friendly side of fabric shows

325
00:09:51,409 --> 00:09:51,690
up.

326
00:09:51,690 --> 00:09:53,769
You don't need to be a database
administrator to add a

327
00:09:53,769 --> 00:09:56,129
relationship or write a simple
measure, because the tools are

328
00:09:56,129 --> 00:09:58,769
visual and the language stays
close to plain English.

329
00:09:58,769 --> 00:10:01,289
The harder part is the thinking
rather than the clicking.

330
00:10:01,289 --> 00:10:04,129
Deciding what revenue should
mean and which rows should count

331
00:10:04,129 --> 00:10:06,769
is a business decision that no
tool makes for you.

332
00:10:06,769 --> 00:10:09,330
Get that thinking right and the
model becomes something the

333
00:10:09,330 --> 00:10:10,570
whole company leans on.

334
00:10:10,570 --> 00:10:12,970
Get it wrong and you've
automated a disagreement.

335
00:10:12,970 --> 00:10:14,769
So let's look at where this
connects to the rest of the

336
00:10:14,769 --> 00:10:17,090
Microsoft tools you already use.

337
00:10:17,090 --> 00:10:18,289
Where it fits with Microsoft
three.

338
00:10:18,289 --> 00:10:20,049
Sixty five and Azure Fabric

339
00:10:20,049 --> 00:10:21,090
doesn't sit off to the side of

340
00:10:21,090 --> 00:10:21,690
Microsoft three.

341
00:10:21,690 --> 00:10:24,049
Sixty five because the whole
point is that your data shows up

342
00:10:24,049 --> 00:10:25,330
where you already work.

343
00:10:25,330 --> 00:10:28,850
Fabric integrates with Excel and
teams, so a number built in a

344
00:10:28,850 --> 00:10:31,409
semantic model can be analyzed
in a spreadsheet or shared in a

345
00:10:31,409 --> 00:10:33,330
channel, without anyone
exporting a file.

346
00:10:33,330 --> 00:10:34,529
For a lot of teams, that's the

347
00:10:34,529 --> 00:10:35,850
moment the model stops feeling

348
00:10:35,850 --> 00:10:37,169
like an IT project and starts

349
00:10:37,169 --> 00:10:38,330
feeling like part of the daily

350
00:10:38,330 --> 00:10:39,090
routine.

351
00:10:39,090 --> 00:10:40,330
Most people will never open

352
00:10:40,330 --> 00:10:41,830
fabric at all, and they'll still

353
00:10:41,830 --> 00:10:42,889
benefit from the work done

354
00:10:42,889 --> 00:10:43,809
inside it.

355
00:10:43,809 --> 00:10:46,690
Underneath one leg speaks the
same language as Azure Data Lake

356
00:10:46,690 --> 00:10:49,250
storage, so existing Azure
services can connect to it.

357
00:10:49,250 --> 00:10:52,169
Azure Synapse Storage Explorer,
and Azure Databricks can all

358
00:10:52,169 --> 00:10:54,889
reach the same data, which means
your data engineers and your

359
00:10:54,889 --> 00:10:57,450
analysts aren't fighting over
two separate copies.

360
00:10:57,450 --> 00:11:00,129
Fabric capacity itself is bought
and built through Azure.

361
00:11:00,129 --> 00:11:03,129
If your company already runs on
Azure, the bill lands somewhere

362
00:11:03,129 --> 00:11:04,690
your finance team recognizes.

363
00:11:04,690 --> 00:11:06,889
Then there's copilot, which
makes the semantic model more

364
00:11:06,889 --> 00:11:08,330
important than it used to be.

365
00:11:08,330 --> 00:11:10,370
Copilot answers questions by
reading the measures,

366
00:11:10,370 --> 00:11:12,690
relationships, and definitions
inside your model.

367
00:11:12,690 --> 00:11:15,250
That means the quality of those
definitions becomes the quality

368
00:11:15,250 --> 00:11:16,049
of the answer.

369
00:11:16,049 --> 00:11:19,049
Ask copilot for sales by region
and it uses the model's version

370
00:11:19,049 --> 00:11:20,649
of sales rather than a guess.

371
00:11:20,649 --> 00:11:21,850
If your measures are well named

372
00:11:21,850 --> 00:11:23,409
and well described, the answers

373
00:11:23,409 --> 00:11:24,129
get sharper.

374
00:11:24,129 --> 00:11:27,409
If they're a mess, the AI passes
the mess along politely.

375
00:11:27,409 --> 00:11:30,169
Microsoft has also started
layering something newer on top

376
00:11:30,169 --> 00:11:31,490
called fabric IQ.

377
00:11:31,490 --> 00:11:33,049
It aims to give AI a shared

378
00:11:33,049 --> 00:11:34,370
business vocabulary across the

379
00:11:34,370 --> 00:11:35,009
platform.

380
00:11:35,009 --> 00:11:38,179
It's still settling in, so treat
the details as moving, but the

381
00:11:38,179 --> 00:11:39,019
direction is clear.

382
00:11:39,019 --> 00:11:42,139
Your model is turning into the
grounding layer for AI, not just

383
00:11:42,139 --> 00:11:43,779
the engine behind a dashboard.

384
00:11:43,779 --> 00:11:44,539
That's the upside.

385
00:11:44,539 --> 00:11:45,259
And it's real.

386
00:11:45,259 --> 00:11:47,820
Now for the honest part, because
none of this comes free and none

387
00:11:47,820 --> 00:11:49,460
of it runs without limits.

388
00:11:49,460 --> 00:11:52,500
The catcher's limits, licensing
and common mistakes start with

389
00:11:52,500 --> 00:11:55,259
licensing because it's the first
thing that trips people up.

390
00:11:55,259 --> 00:11:58,179
Fabric runs on capacity, which
is a pool of computing power.

391
00:11:58,179 --> 00:11:59,659
And on top of that, you need a

392
00:11:59,659 --> 00:12:00,980
per user license to do most of

393
00:12:00,980 --> 00:12:01,419
the work.

394
00:12:01,419 --> 00:12:04,379
There's a twist that surprises
teams on smaller capacities.

395
00:12:04,379 --> 00:12:05,700
Anyone who just wants to view a

396
00:12:05,700 --> 00:12:06,700
report still needs a paid

397
00:12:06,700 --> 00:12:07,379
license.

398
00:12:07,379 --> 00:12:11,259
Free viewing only unlocks at x64
and above, which is a big jump

399
00:12:11,259 --> 00:12:12,820
from the entry level options.

400
00:12:12,820 --> 00:12:15,419
Direct Lake has its own rule,
and it's stricter.

401
00:12:15,419 --> 00:12:18,500
It needs a fabric capacity or a
premium capacity behind it, so

402
00:12:18,500 --> 00:12:21,460
the free per user license on its
own won't run it.

403
00:12:21,460 --> 00:12:24,500
Import and direct query models
are more forgiving and work on

404
00:12:24,500 --> 00:12:25,980
any license, including free.

405
00:12:25,980 --> 00:12:26,820
That's worth remembering.

406
00:12:26,820 --> 00:12:28,179
If you're learning on a trial.

407
00:12:28,179 --> 00:12:29,580
Then come the capacity limits.

408
00:12:29,580 --> 00:12:31,980
And these are the numbers people
hit in production.

409
00:12:31,980 --> 00:12:34,820
Every capacity size has
guardrails for rows per table,

410
00:12:34,820 --> 00:12:38,820
files per table, and memory on
the smallest capacities from F

411
00:12:38,820 --> 00:12:39,860
two to F eight.

412
00:12:39,860 --> 00:12:42,539
You get around three hundred
million rows per table, one

413
00:12:42,539 --> 00:12:45,700
thousand parquet files and three
gigabytes of memory.

414
00:12:45,700 --> 00:12:46,899
Step up to F sixty four and

415
00:12:46,899 --> 00:12:48,000
those limits jump to one and a

416
00:12:48,000 --> 00:12:49,500
half billion rows, five thousand

417
00:12:49,500 --> 00:12:50,360
files, and twenty five

418
00:12:50,360 --> 00:12:50,940
gigabytes.

419
00:12:50,940 --> 00:12:52,500
What happens when you blow past
them?

420
00:12:52,500 --> 00:12:55,460
Depends on which direct lake
you're using on direct lake.

421
00:12:55,460 --> 00:12:58,860
On one lake, the refresh simply
fails, and the model can't be

422
00:12:58,860 --> 00:13:01,700
queried until someone fixes the
underlying tables on the older

423
00:13:01,700 --> 00:13:02,899
direct lake on SQL.

424
00:13:02,899 --> 00:13:05,820
The model falls back to direct
query, so reports keep working

425
00:13:05,820 --> 00:13:06,779
but get slower.

426
00:13:06,779 --> 00:13:08,059
Neither one is a problem you can

427
00:13:08,059 --> 00:13:09,419
ignore because the fix is real

428
00:13:09,419 --> 00:13:10,179
work.

429
00:13:10,179 --> 00:13:11,500
The mistakes people make tend to

430
00:13:11,500 --> 00:13:12,419
repeat, and they're worth

431
00:13:12,419 --> 00:13:13,019
naming.

432
00:13:13,019 --> 00:13:15,899
Teams leave behind old default
models that nobody owns.

433
00:13:15,899 --> 00:13:18,820
They bury heavy data preparation
inside the model, where it's

434
00:13:18,820 --> 00:13:20,019
invisible to everyone else.

435
00:13:20,019 --> 00:13:21,340
They deploy a model and forget

436
00:13:21,340 --> 00:13:22,820
to refresh it so it serves stale

437
00:13:22,820 --> 00:13:23,620
data.

438
00:13:23,620 --> 00:13:24,899
They let one model grow per

439
00:13:24,899 --> 00:13:26,580
report instead of sharing one,

440
00:13:26,580 --> 00:13:27,460
until there are twenty

441
00:13:27,460 --> 00:13:29,019
definitions of the same metric,

442
00:13:29,019 --> 00:13:30,419
and they forget that sharing a

443
00:13:30,419 --> 00:13:32,340
report doesn't automatically

444
00:13:32,340 --> 00:13:33,740
grant access to the data

445
00:13:33,740 --> 00:13:35,320
underneath, which produces

446
00:13:35,320 --> 00:13:36,759
access errors that are hard to

447
00:13:36,759 --> 00:13:37,460
trace.

448
00:13:37,460 --> 00:13:40,379
Copilot has limits to it,
answers in English only, and it

449
00:13:40,379 --> 00:13:42,340
can't do forecasting or spot
anomalies for you.

450
00:13:42,340 --> 00:13:44,460
Yet it's a helper, not an
oracle.

451
00:13:44,460 --> 00:13:45,820
None of this should scare you
off.

452
00:13:45,820 --> 00:13:47,179
A fabric trial gives you sixty

453
00:13:47,179 --> 00:13:48,580
days to experiment, and that's

454
00:13:48,580 --> 00:13:49,820
usually enough to see whether

455
00:13:49,820 --> 00:13:51,179
any of it fits the way your team

456
00:13:51,179 --> 00:13:52,259
works.

457
00:13:52,259 --> 00:13:53,940
The recap and one question for
you.

458
00:13:53,940 --> 00:13:55,419
So let's bring it all together.

459
00:13:55,419 --> 00:13:57,740
A semantic model holds the
business definitions for your

460
00:13:57,740 --> 00:14:01,299
data, and it's the layer every
report reads from fabric, feeds

461
00:14:01,299 --> 00:14:03,899
it straight from one lake using
Direct Lake, which reads just

462
00:14:03,899 --> 00:14:05,980
the files it needs without
copying them all.

463
00:14:05,980 --> 00:14:08,899
Framing keeps it current without
an overnight refresh.

464
00:14:08,899 --> 00:14:10,179
The catches are real and they

465
00:14:10,179 --> 00:14:11,740
live in licensing, in capacity

466
00:14:11,740 --> 00:14:13,159
limits, and in the everyday

467
00:14:13,159 --> 00:14:14,620
discipline of naming things well

468
00:14:14,620 --> 00:14:15,899
and sharing one model instead of

469
00:14:15,899 --> 00:14:16,700
twenty.

470
00:14:16,700 --> 00:14:19,860
None of it is magic, and it
isn't beyond a curious beginner.

471
00:14:19,860 --> 00:14:21,220
I'll leave you with one
question.

472
00:14:21,220 --> 00:14:23,840
If you could define one number
so that everyone in your company

473
00:14:23,840 --> 00:14:26,340
finally agreed on it, which
number would it be?

474
00:14:26,340 --> 00:14:27,620
If this made the topic feel less

475
00:14:27,620 --> 00:14:30,000
intimidating, subscribe to M365

476
00:14:30,000 --> 00:14:31,419
FM to master the tools and the

477
00:14:31,419 --> 00:14:32,740
mindset shaping the future of

478
00:14:32,740 --> 00:14:33,340
work.
