1
00:00:00,024 --> 00:00:01,854
One of the things that we've been seeing inside ElastiCache is

2
00:00:01,854 --> 00:00:04,244
we're trying to contribute more of our, you know, performance

3
00:00:04,254 --> 00:00:06,724
features, our efficiency features back into open source, because

4
00:00:06,724 --> 00:00:10,073
we think that helps us make the overall system more stable

5
00:00:15,282 --> 00:00:16,872
Welcome to Screaming in the Cloud.

6
00:00:17,072 --> 00:00:21,522
I'm Corey Quinn, and it is a pleasure for me once again to speak with

7
00:00:21,532 --> 00:00:27,302
Madelyn Olson, AWS principal engineer and core maintainer of Valkey.

8
00:00:27,381 --> 00:00:28,561
Madelyn, how have you been?

9
00:00:29,002 --> 00:00:30,152
Oh, I've been delightful.

10
00:00:30,162 --> 00:00:32,002
Thank you so much for having me back.

11
00:00:32,231 --> 00:00:36,151
Well, I, frankly, I'm astounded and grateful that you agreed

12
00:00:36,151 --> 00:00:39,091
to come back on the show because we gave a talk at re:Invent,

13
00:00:39,121 --> 00:00:41,892
and usually that's when the knives come out afterwards.

14
00:00:41,892 --> 00:00:45,421
Like, I can't believe you said that on stage, but apparently people liked it.

15
00:00:45,551 --> 00:00:49,141
One of my favorite things about our talk was that everyone came up afterwards,

16
00:00:49,141 --> 00:00:51,862
and everyone wanted to talk to you, which is just so… It's so nice.

17
00:00:51,871 --> 00:00:53,291
Everyone usually wants to talk to me.

18
00:00:53,301 --> 00:00:53,751
It's okay.

19
00:00:53,752 --> 00:00:55,501
I can be an introvert and just walk away.

20
00:00:55,891 --> 00:00:58,491
It has its advantages and disadvantages.

21
00:00:58,502 --> 00:01:00,852
Sometimes people wanna say something kind.

22
00:01:01,191 --> 00:01:03,571
Other times they want… Like, the thing that drives me nuts is right

23
00:01:03,571 --> 00:01:06,401
after you step off of a stage, and people wanna give you harsh feedback

24
00:01:06,401 --> 00:01:10,471
on the talk, which is helpful as a speaker, don't get me wrong, but give

25
00:01:10,471 --> 00:01:13,631
me 10 minutes first to wind down from the high of having given the talk

26
00:01:13,821 --> 00:01:17,331
or, you know, ask me to sign their chest or get a selfie or smell my hair.

27
00:01:17,331 --> 00:01:19,921
Eh, okay, it gets a little interesting, but we run with it.

28
00:01:20,301 --> 00:01:20,972
Oh, it's true.

29
00:01:20,972 --> 00:01:24,662
I have, you know, have to deal with my small share of fame being

30
00:01:24,662 --> 00:01:27,811
one of the Valkey maintainers, but I'm sure you get much more of it.

31
00:01:28,512 --> 00:01:30,401
Hey, you're the one that got me into the Redis party.

32
00:01:30,421 --> 00:01:30,911
That was fun.

33
00:01:31,241 --> 00:01:33,701
Uh, I think they would have let anyone in.

34
00:01:33,871 --> 00:01:37,181
Entirely pos- That's the beautiful part about being open source,

35
00:01:37,182 --> 00:01:39,922
then not being open source, then barely being open source again.

36
00:01:40,690 --> 00:01:43,650
So I have some questions for you because

37
00:01:43,740 --> 00:01:46,270
so much has changed in the, what is it now?

38
00:01:46,280 --> 00:01:50,800
Six months since re:Invent, give or take, where back then a lot of people

39
00:01:50,800 --> 00:01:56,300
were throwing AI pull requests to open source projects, and it was slop.

40
00:01:56,309 --> 00:01:58,850
I mean, no one would pass up a sl opportunity to wind up

41
00:01:58,860 --> 00:02:01,670
YOLO coding something at, that, that doesn't freaking work.

42
00:02:02,020 --> 00:02:06,409
And at some point in that interim, it feels like something

43
00:02:06,410 --> 00:02:11,480
has shifted, namely the AI-generated code started being good.

44
00:02:12,139 --> 00:02:15,960
What's it like as the maintainer of a project people actually care about

45
00:02:15,960 --> 00:02:21,110
and contribute to, to, to be, I guess, facing that, that rising tide?

46
00:02:21,649 --> 00:02:24,740
So we definitely saw the increase of pull requests

47
00:02:24,779 --> 00:02:27,519
coming a little bit late last year, but by and large they

48
00:02:27,520 --> 00:02:30,529
were, as you said, the PRs were not very good quality.

49
00:02:30,529 --> 00:02:31,889
They had obvious mistakes.

50
00:02:31,889 --> 00:02:36,479
They also often… They were often trying to find issues inside the project.

51
00:02:36,479 --> 00:02:38,570
They'd, like, go and search through the issues and

52
00:02:38,570 --> 00:02:42,439
find something, and they would kind of just send the AI

53
00:02:42,439 --> 00:02:44,700
agent at it with a one-shot being like, "Go fix this."

54
00:02:44,729 --> 00:02:46,979
And the result were what you would expect.

55
00:02:46,990 --> 00:02:47,989
They were low quality.

56
00:02:47,990 --> 00:02:49,230
They didn't make very much sense, so we

57
00:02:49,230 --> 00:02:50,970
kind of quickly were starting to close them.

58
00:02:51,240 --> 00:02:53,979
We definitely noticed about three to four months ago

59
00:02:53,979 --> 00:02:58,019
that the average PR quality was both increasing a lot.

60
00:02:58,529 --> 00:03:02,100
Basically, average engineers were able to point their, you

61
00:03:02,100 --> 00:03:06,249
know, Claude Code or Kira or some other types of harnesses at,

62
00:03:06,289 --> 00:03:09,979
with the foundational, like, Opus models with, from Anthropic.

63
00:03:10,289 --> 00:03:12,109
They were able to point them at these issues, and

64
00:03:12,109 --> 00:03:14,360
they were actually producing semi-decent code.

65
00:03:14,719 --> 00:03:18,130
And that started causing a lot of pain on a lot of open source

66
00:03:18,189 --> 00:03:21,859
maintainers because that took a huge amount of effort to sort of

67
00:03:21,859 --> 00:03:25,599
review and think through, 'cause the AI models still aren't great at,

68
00:03:25,669 --> 00:03:29,170
like, making sure the code is adhering to, like, sort of the ethos

69
00:03:29,170 --> 00:03:32,189
of the project, but it's much better at being technically correct

70
00:03:32,558 --> 00:03:34,178
Which is, of course, the best kind of correct.

71
00:03:34,188 --> 00:03:35,338
Of course, it's the best kind of correct.

72
00:03:35,368 --> 00:03:38,678
So yes, it doesn't crash, but it also is gonna be hard to maintain.

73
00:03:39,227 --> 00:03:41,908
AI models still love generating code, so even if there's

74
00:03:41,908 --> 00:03:44,338
a perfectly good function you could rewrite a little bit,

75
00:03:44,347 --> 00:03:46,567
they're gonna totally just go off and build their own.

76
00:03:46,867 --> 00:03:50,018
My favorite thing with like Claude Code is, so the whole Valkey

77
00:03:50,018 --> 00:03:53,857
project is written in Tcl, a archaic language, but the, the…

78
00:03:53,867 --> 00:03:56,097
Sorry, the infrastructure, the testing is all written in Tcl.

79
00:03:56,478 --> 00:04:00,368
The AI loves to just write its own Tcl frameworks every single time if you don't

80
00:04:00,368 --> 00:04:03,817
like hold it to the fire and you're like, "No, you have to use the framework

81
00:04:03,817 --> 00:04:07,307
we built." It loves writing new Tcl tests, but that's not very maintainable.

82
00:04:07,848 --> 00:04:10,257
So we definitely noticed in the last couple of months that they actually started

83
00:04:10,258 --> 00:04:14,048
doing a really good job of like, you know, kind of getting almost the way there.

84
00:04:14,427 --> 00:04:18,367
And so we've seen about a 50 to 60% increase of actually decent

85
00:04:18,378 --> 00:04:20,737
pull requests getting opened in the last couple of months.

86
00:04:21,187 --> 00:04:23,587
And that's sort of been the big thing we've been trying to deal with,

87
00:04:23,587 --> 00:04:27,557
like how do we as a open source project deal with this, these increases in

88
00:04:27,558 --> 00:04:31,267
contributions that w- take a lot of maintainer time to actually think through?

89
00:04:31,697 --> 00:04:33,627
'Cause they don't have bugs, they just kind

90
00:04:33,628 --> 00:04:36,627
of aren't great, but they're not bad either.

91
00:04:36,627 --> 00:04:40,397
I, I have found that one of the problems with AI-generated code

92
00:04:40,407 --> 00:04:44,517
historically is that people continue to make the same fundamental

93
00:04:44,558 --> 00:04:48,188
error about AI that they have been making for almost four years

94
00:04:48,188 --> 00:04:51,428
now, which is, "Hey, this thing talks like a middle manager.

95
00:04:51,558 --> 00:04:54,548
Therefore, it must be self-aware," instead of the proper

96
00:04:54,548 --> 00:04:57,467
conclusion, which is that middle managers are absolutely not.

97
00:04:57,708 --> 00:05:00,338
Where people hate AI slop, in my experience, from

98
00:05:00,338 --> 00:05:03,337
a, from a narrative perspective, has been when it

99
00:05:03,378 --> 00:05:08,407
presents as mediocre, as milk toast corporate speak.

100
00:05:08,628 --> 00:05:12,238
If you throw whimsy into it, people find it delightful.

101
00:05:12,387 --> 00:05:16,247
Uh, my AI-generated code commits are no better than anyone

102
00:05:16,247 --> 00:05:19,918
else's, but I will say that my commits are the ones that have

103
00:05:19,918 --> 00:05:23,187
conspiracy theories about the code in the commit message.

104
00:05:23,617 --> 00:05:25,708
'Cause if you're gonna put out slop, at least be funny about it.

105
00:05:26,304 --> 00:05:27,244
Oh no, for sure.

106
00:05:27,264 --> 00:05:30,514
People hate the, the, the… There's like that speak that especially I

107
00:05:30,514 --> 00:05:33,404
think, you know, ChatGPT got known for that, you know, every time you're

108
00:05:33,404 --> 00:05:35,754
like, "Hey, that's just wrong," it's like, "Oh, you're absolutely right.

109
00:05:35,794 --> 00:05:38,503
I just didn't do what you told me to do." But if you throw a little

110
00:05:38,504 --> 00:05:41,333
more whimsy in it, it is, it can give much funnier responses.

111
00:05:41,593 --> 00:05:45,274
It's like, "Yes, I am deliberately trying to sabotage you." It's a lot more fun.

112
00:05:45,754 --> 00:05:47,903
Everyone has been talking about Mythos

113
00:05:47,934 --> 00:05:50,613
because it's a brilliant marketing campaign.

114
00:05:50,734 --> 00:05:52,304
We built this amazing model.

115
00:05:52,383 --> 00:05:54,284
Oh, too scary to release.

116
00:05:54,284 --> 00:05:56,463
Special people can have it, but not you.

117
00:05:57,213 --> 00:06:01,814
And it, it's become this weird, uh, I have something super special.

118
00:06:01,844 --> 00:06:03,294
Can I… Can, can other people see it?

119
00:06:03,783 --> 00:06:04,013
No.

120
00:06:04,303 --> 00:06:09,003
Then you had, uh, GPT-55 come out, which apparently is similar, only

121
00:06:09,053 --> 00:06:14,084
other folks get to use that one, and it has definitely re- raised the bar.

122
00:06:14,274 --> 00:06:17,714
I will say that I have the numbers on this, which I'm sure your

123
00:06:17,754 --> 00:06:21,593
friends in AWS PR are gonna love me for, but I've done visualizations

124
00:06:21,623 --> 00:06:25,953
on the number of CVEs that AWS has published on an annual basis for

125
00:06:25,953 --> 00:06:30,263
the last five years, and we are gonna cross a threshold this month.

126
00:06:30,603 --> 00:06:32,853
For the record, we are now at the beginning of May

127
00:06:33,582 --> 00:06:33,962
Yeah.

128
00:06:34,182 --> 00:06:38,272
So AI does have a great ability to go and force multiply in

129
00:06:38,272 --> 00:06:41,661
a way that individuals can't, and I think where I was kind of

130
00:06:41,712 --> 00:06:44,392
talking about, like, the harnesses around the MLs have gotten

131
00:06:44,392 --> 00:06:47,071
a lot better, and, like, I don't have access to Meet Those.

132
00:06:47,431 --> 00:06:49,911
I don't know anything that I can share

133
00:06:49,912 --> 00:06:51,831
publicly about, you know, access to Meet Those.

134
00:06:52,141 --> 00:06:55,591
But inside the Valkey project, we've actually been quite successful at just kind

135
00:06:55,592 --> 00:06:59,931
of, you know, taking, you know, the frontier Opus models and being like, "Hey,

136
00:06:59,931 --> 00:07:03,931
just go read through this code very deeply. Try to come up with hypotheses about

137
00:07:03,931 --> 00:07:07,872
what might be exploitable," and then trying to… Then do- go try to fix it.

138
00:07:08,341 --> 00:07:10,112
We've been actually generating a lot of PRs

139
00:07:10,112 --> 00:07:12,101
and fixes recently inside the Valkey project.

140
00:07:12,111 --> 00:07:15,852
I think the total's up to, like, 25 or 26 in the last month or so of not

141
00:07:15,852 --> 00:07:19,892
CVE caliber bugs, but, like, of real bugs, like memory leaks and unintended

142
00:07:19,892 --> 00:07:24,841
server asserts by just, you know, trying to, you know, use a lot of tokens.

143
00:07:25,161 --> 00:07:28,671
You know, I- it's nice that at Amazon I'm able to burn kind of as many

144
00:07:28,672 --> 00:07:32,351
tokens as I want sort of doing these deep evaluations of the code base,

145
00:07:32,821 --> 00:07:36,182
and we've been able to do… use that to find bugs, and I'm sure Meet

146
00:07:36,182 --> 00:07:40,322
Those is the same, and I, I think AI will really help us harden the code.

147
00:07:40,491 --> 00:07:41,181
Oh, it has.

148
00:07:41,181 --> 00:07:43,721
I mean, my, my code is no great shakes, but my internal system

149
00:07:43,732 --> 00:07:47,071
for writing the newsletter is open to the public, and I finally

150
00:07:47,071 --> 00:07:49,841
decided that I had some spare cycles left in a session, and, all

151
00:07:49,841 --> 00:07:51,961
right, I'm gonna go ahead and do a security audit of this thing.

152
00:07:52,241 --> 00:07:55,312
Uh, so the passcode to get in is a UUID.

153
00:07:55,552 --> 00:07:58,871
It turns out that anything of the appropriate length would suffice.

154
00:07:59,451 --> 00:08:01,901
Like, okay, that, that's not terrific.

155
00:08:02,211 --> 00:08:04,731
Uh, again, the, the blast radius is somewhat

156
00:08:04,741 --> 00:08:07,302
minimal on this, but still not a great failure mode.

157
00:08:07,561 --> 00:08:09,381
And then I found a bunch of other stuff, which, where it

158
00:08:09,381 --> 00:08:11,691
gets into the stupid things I don't really care about.

159
00:08:11,701 --> 00:08:14,571
Like, hey, because I'm the only user that has access

160
00:08:14,571 --> 00:08:17,661
to this, but I could potentially prompt inject myself.

161
00:08:17,692 --> 00:08:19,081
It… Great.

162
00:08:19,082 --> 00:08:20,972
Or not prompt inject, but you know what I'm talking about.

163
00:08:20,972 --> 00:08:23,732
I could effectively… I could do SQL injection against

164
00:08:23,741 --> 00:08:26,201
my own code base where I have admin rights anyway.

165
00:08:26,591 --> 00:08:27,591
That's not the threat.

166
00:08:27,762 --> 00:08:30,131
Randos coming in off the internet, more of a threat.

167
00:08:30,582 --> 00:08:30,921
No, yeah.

168
00:08:30,951 --> 00:08:33,372
And one of the great things about AI is it scales really well.

169
00:08:33,391 --> 00:08:37,761
So you can use it and go and hunt down all of these things at the same time.

170
00:08:38,251 --> 00:08:41,261
Obviously, the ones, th- those threats that you talked about, like, the

171
00:08:41,261 --> 00:08:44,062
false positive rate's still the real problem that we're dealing with.

172
00:08:44,421 --> 00:08:48,172
I mentioned, you know, we were able to find, like, 25 or so bugs inside Valkey.

173
00:08:48,552 --> 00:08:50,921
That's nothing compared to how many false positives it generated.

174
00:08:50,931 --> 00:08:52,991
It generated, like, 150 false positives.

175
00:08:52,992 --> 00:08:55,821
It was like, oh, if, you know, this variable somehow got

176
00:08:55,821 --> 00:08:58,981
modified, then you could, you know, cause a remote code execution.

177
00:08:58,981 --> 00:09:00,021
I'm like, yeah, but that can't…

178
00:09:00,031 --> 00:09:01,452
You can't modify the code that way.

179
00:09:01,461 --> 00:09:03,382
That's not how this, this works.

180
00:09:03,461 --> 00:09:06,041
But, you know, my expectation is they're gonna keep getting

181
00:09:06,041 --> 00:09:09,332
better, and we can keep using these tools to basically

182
00:09:09,332 --> 00:09:12,031
harden the systems that we rely on, which I think is cool

183
00:09:12,590 --> 00:09:16,220
I, I will say it does feel like it's gotten harder to contribute to open

184
00:09:16,220 --> 00:09:20,730
source given the proliferation of AI, which, which is counterintuitive because

185
00:09:20,730 --> 00:09:24,800
it has never been easier for me to ha- find something annoying in a project

186
00:09:25,229 --> 00:09:28,980
in a la- written in a language that I don't know, you know, English, and

187
00:09:29,039 --> 00:09:34,090
then I can basically bully the AI into writing a pull request against it.

188
00:09:34,289 --> 00:09:38,879
But I worry that if I do that, then I'm actually part of the problem because

189
00:09:38,930 --> 00:09:42,509
no matter how much I work on that and tweak it and, and get it to do the thing

190
00:09:42,510 --> 00:09:44,960
and add tests and do all, and jump through all the hoops you're supposed to,

191
00:09:45,459 --> 00:09:48,159
there's very little signal, you know, other than the conspiracy theories,

192
00:09:48,279 --> 00:09:52,520
to differentiate this from any other slopportunity people are jumping on.

193
00:09:53,139 --> 00:09:56,979
Yeah, and that's, that's a very fair and valid concern, and I don't know

194
00:09:56,979 --> 00:10:00,360
if we've really figured this out in Valkey yet, but one of the things that

195
00:10:00,360 --> 00:10:03,109
I know I've seen a lot with other maintainers is they've been a relatively

196
00:10:03,110 --> 00:10:06,540
slow embrace of AI 'cause they've seen these, you know, waves of slop.

197
00:10:06,549 --> 00:10:08,200
They, they don't like dealing with them.

198
00:10:08,200 --> 00:10:09,449
They often just close them.

199
00:10:09,849 --> 00:10:11,829
But I have seen more of a shift recently.

200
00:10:11,860 --> 00:10:14,660
The various maintainers in Valkey, I work

201
00:10:14,660 --> 00:10:16,999
with a maintainer from Google named Jacob.

202
00:10:17,009 --> 00:10:19,819
He started using AI to do a lot more reviewing of the code.

203
00:10:20,189 --> 00:10:22,269
We also have an engineer from Erickson who's been using

204
00:10:22,279 --> 00:10:25,389
Kiro, uh, along with Claude to do more reviews of the code.

205
00:10:25,869 --> 00:10:28,609
And so I think right now it definitely feels like there's this

206
00:10:28,619 --> 00:10:31,959
tension, but I think that we're kind of all trying to figure out

207
00:10:32,079 --> 00:10:35,609
what the future of open source development looks like with AI.

208
00:10:36,139 --> 00:10:40,259
And I still appreciate when someone is trying to fix a bug.

209
00:10:40,889 --> 00:10:42,789
You don't, you're not required to submit a p- PR.

210
00:10:42,799 --> 00:10:44,790
You can also just open an issue, and one of the nice things

211
00:10:44,790 --> 00:10:47,879
about generative code is, you know, I can also just go try to fix

212
00:10:47,880 --> 00:10:51,470
your bug with my own GenAI, so you don't have to deal with it.

213
00:10:52,109 --> 00:10:54,659
But I still appreciate when people try to open PRs.

214
00:10:54,669 --> 00:10:58,309
So I definitely don't want people to start feeling like that's

215
00:10:59,328 --> 00:11:01,588
Like they're contributing to the problem by trying to help.

216
00:11:01,658 --> 00:11:03,078
Most people are good intentioned.

217
00:11:03,458 --> 00:11:06,468
The people I really don't like are the people who like use

218
00:11:06,468 --> 00:11:09,677
like OpenClaw and like they just point at the Valkey project

219
00:11:09,677 --> 00:11:11,318
and they're like, "Yeah, just go try to fix all the issues."

220
00:11:11,627 --> 00:11:13,688
Like generally those issues are there 'cause we'd like

221
00:11:13,688 --> 00:11:16,307
people to learn about the project, get involved in the

222
00:11:16,307 --> 00:11:18,798
project, or maybe they're issues that aren't very important.

223
00:11:19,307 --> 00:11:22,387
So the people that just use AI to, you know, brute force and try to solve

224
00:11:22,387 --> 00:11:25,357
lots of problems are the people that are really causing the problems, not

225
00:11:25,358 --> 00:11:28,607
the, not the average individual who's trying to make the project better.

226
00:11:28,907 --> 00:11:30,047
So I wanna talk a bit about it.

227
00:11:30,087 --> 00:11:34,528
It's now two years since the launch of Valkey, and in

228
00:11:34,568 --> 00:11:37,768
some ways it has succeeded from the customer perspective,

229
00:11:38,008 --> 00:11:41,178
uh, the biggest concerns that folks had with Redis.

230
00:11:41,397 --> 00:11:42,848
One was the attempted rug pull.

231
00:11:42,857 --> 00:11:43,187
Great.

232
00:11:43,237 --> 00:11:43,647
Awesome.

233
00:11:43,657 --> 00:11:46,687
The, the community made a definitive decision to fork, which is

234
00:11:46,688 --> 00:11:50,538
great, and that in turn unblocked a lot of features that, not to sound

235
00:11:50,538 --> 00:11:53,968
uncharitable, it felt like Redis was intentionally holding back as a form

236
00:11:53,968 --> 00:11:57,597
of business model protection, and now the open source version is awesome.

237
00:11:57,947 --> 00:11:59,527
How has it been from the other side of it?

238
00:11:59,567 --> 00:12:02,387
Because I'm, I'm just the customer, and to be direct, I'll

239
00:12:02,397 --> 00:12:06,967
use whatever, uh, whatever version of this that my AI agent

240
00:12:07,007 --> 00:12:09,828
picks most of the time, but you see it very differently.

241
00:12:10,487 --> 00:12:10,797
Yeah.

242
00:12:10,807 --> 00:12:13,497
So the first year after the fork, there was definitely

243
00:12:13,498 --> 00:12:15,907
this big open question of, "Hey, will Valkey survive?"

244
00:12:15,937 --> 00:12:16,247
Right?

245
00:12:16,257 --> 00:12:17,597
Most forks don't work out.

246
00:12:17,598 --> 00:12:18,877
Most forks fizzle out.

247
00:12:19,574 --> 00:12:22,064
And one of the great things we did see about Valkey is it did end up

248
00:12:22,064 --> 00:12:25,354
surviving, in part because it was sponsored by a large number of organizations.

249
00:12:25,414 --> 00:12:27,654
We had a very diverse community that was helping build it.

250
00:12:28,324 --> 00:12:30,413
And we even s- did see some validation when Redis

251
00:12:30,414 --> 00:12:32,814
ended up moving back to AGPL so quickly, right?

252
00:12:32,814 --> 00:12:35,184
They moved from a very permissive license

253
00:12:35,184 --> 00:12:37,983
to, to proprietary license back to AGPL.

254
00:12:38,534 --> 00:12:41,854
And the users from our project that saw that kind of saw that as

255
00:12:41,864 --> 00:12:45,263
validation that, hey, like, the Valkey project is… It's real.

256
00:12:45,284 --> 00:12:46,754
People are thinking about using it.

257
00:12:47,264 --> 00:12:49,714
And over the last year, we've seen basically more and more

258
00:12:49,714 --> 00:12:52,573
adoption of the project, 'cause people are seeing it stick around.

259
00:12:52,574 --> 00:12:53,974
They're seeing it build functionality

260
00:12:53,974 --> 00:12:55,843
that, you know, wasn't showing up in Redis.

261
00:12:56,183 --> 00:13:00,013
The things you mentioned, you know, like we built LDAP support, which was

262
00:13:00,234 --> 00:13:04,013
a long time a proprietary feature of Redis, and, you know, that helps build

263
00:13:04,014 --> 00:13:07,584
a lot of confidence that we're willing to build stuff that end users want.

264
00:13:07,784 --> 00:13:10,774
I was even talking with a, a financial company this

265
00:13:10,774 --> 00:13:13,683
morning, and they're like, "Hey, we're now all in on Valkey.

266
00:13:13,714 --> 00:13:14,843
We'd like to talk about it.

267
00:13:14,854 --> 00:13:17,704
We have some v- very esoteric features that

268
00:13:17,714 --> 00:13:19,843
we were never able to get merged into Redis.

269
00:13:19,874 --> 00:13:22,443
Like, can you help us merge it?" They're doing some… They

270
00:13:22,444 --> 00:13:25,854
basically want, like, a chaos testing, like API and Valkey,

271
00:13:25,854 --> 00:13:28,683
so that they can better verify their availability guarantees

272
00:13:28,684 --> 00:13:31,803
of the pro- uh, of their service when the cache goes down.

273
00:13:32,523 --> 00:13:33,574
And those are really cool things.

274
00:13:33,574 --> 00:13:36,383
And so, like, you know, the power of Valkey right now is still its community.

275
00:13:36,873 --> 00:13:39,043
And as I said, we're seeing more and more adoption.

276
00:13:39,053 --> 00:13:40,993
We're still seeing… We recently cro- crossed 100

277
00:13:41,833 --> 00:13:44,073
million container pulls of the product overall.

278
00:13:44,114 --> 00:13:46,883
If you kinda look at the graph, it's been, like, a nice upward

279
00:13:47,193 --> 00:13:50,993
exponential graph as, you know, we see more and more adoption.

280
00:13:51,284 --> 00:13:51,613
Oh, yeah.

281
00:13:51,644 --> 00:13:54,833
And as I mentioned in our talk that we gave, which I'll throw

282
00:13:54,833 --> 00:13:59,833
a link to in the show notes, uh, that there was a… They were

283
00:13:59,863 --> 00:14:03,943
pulling a number of, of upstream commits from Valkey into Redis.

284
00:14:04,313 --> 00:14:07,763
It has become the new… Uh, actually, Valkey has become the

285
00:14:07,763 --> 00:14:12,003
new Redis upstream, which is a terrific example of success.

286
00:14:12,233 --> 00:14:15,283
So for those who are not able to look at the show notes,

287
00:14:15,284 --> 00:14:19,414
this is OPN 309, titled appropriately, Disagree in Commits.

288
00:14:19,414 --> 00:14:22,543
And we'll throw a link to that into the show notes because of course we will.

289
00:14:22,964 --> 00:14:23,394
And yeah.

290
00:14:23,423 --> 00:14:26,353
And, you know, Redis is still taking commits from us, uh, which is great.

291
00:14:26,473 --> 00:14:27,384
We love to see it.

292
00:14:27,473 --> 00:14:30,273
You know, it's, it is definitely affirmation that what we're building is

293
00:14:30,293 --> 00:14:34,303
valuable to the end users, and we kind of believe in our long-term ability

294
00:14:34,304 --> 00:14:38,563
to continue to, you know, build things that our end users find valuable

295
00:14:39,244 --> 00:14:39,874
And that's important.

296
00:14:39,874 --> 00:14:43,224
It, it comes down to the customer obsession approach.

297
00:14:43,664 --> 00:14:48,554
So I, I guess the, the counterpoint that I, uh, that I want to bring

298
00:14:48,554 --> 00:14:52,004
up here, the sort of the elephant in the room, such as it is, when

299
00:14:52,044 --> 00:14:56,063
everyone can have the AI service wind up writing and submitting

300
00:14:56,064 --> 00:14:59,373
code with less and less human effort, uh, several things happen.

301
00:14:59,374 --> 00:15:02,803
One of them is GitHub falls down more often than grandma.

302
00:15:03,554 --> 00:15:08,903
I get it, scale is hard, and yet the other piece of it is that you wind

303
00:15:08,913 --> 00:15:14,054
up with a tidal wave of, I'm going to be uncharitable for a second, crap.

304
00:15:14,453 --> 00:15:16,003
How do you, how do you fight that?

305
00:15:16,034 --> 00:15:20,044
How do you push back against that upswell of massive nonsense?

306
00:15:20,664 --> 00:15:24,103
You know, one of the things that's great about AI is its non-determinism, right?

307
00:15:24,103 --> 00:15:26,994
If you ask it to build a thing, it'll build it 10 different ways,

308
00:15:27,083 --> 00:15:29,744
which is counter to what we want in what you just described, right?

309
00:15:29,744 --> 00:15:30,853
We want determinism.

310
00:15:30,863 --> 00:15:35,033
We want things that stay up, stay available, and, like, don't break, right?

311
00:15:35,113 --> 00:15:38,053
Which is basically things like writing tests, writing automation.

312
00:15:38,543 --> 00:15:40,663
Like, one of the things that's important that, you know, I was really

313
00:15:40,663 --> 00:15:44,153
internalized while I've worked at AWS is, like, we have huge suites

314
00:15:44,153 --> 00:15:47,584
of automated testing that does chaos testing, that does, you know,

315
00:15:47,584 --> 00:15:50,914
regression testing and, like, that's something that we need to be

316
00:15:50,914 --> 00:15:55,463
using AI to both write tests for, to do ver- verification against.

317
00:15:55,813 --> 00:15:59,464
Like, one of the things I mentioned is Redis is able to pull commits from

318
00:15:59,464 --> 00:16:02,713
us, but we're not able to pull commits from them for licensing reasons.

319
00:16:02,964 --> 00:16:05,743
And one of the things we built is some tooling around basically every commit

320
00:16:05,743 --> 00:16:10,763
that gets opened to Valkey, we check to see if there's any chance that that

321
00:16:10,773 --> 00:16:13,943
commit might have originated from Redis by comparing it against hashes of the

322
00:16:13,943 --> 00:16:18,401
code base And that's important for us 'cause, like, we really d- wanna make sure

323
00:16:18,401 --> 00:16:22,692
we don't accidentally pull a commit 'cause it's very painful to unwind that.

324
00:16:23,382 --> 00:16:25,172
And then also stuff, you know, like we've been using

325
00:16:25,172 --> 00:16:27,021
AI to, you know, write a lot of regression testing.

326
00:16:27,032 --> 00:16:30,451
We built… We used AI tooling to build fuzzing against the Valkey

327
00:16:30,452 --> 00:16:34,202
system to basically test how, how it behaves when various nodes fails.

328
00:16:34,742 --> 00:16:39,401
And so, yes, on one side there is this rise of just non-deterministic generative

329
00:16:39,432 --> 00:16:44,551
code, but we can also use that ability to generate deterministic code that

330
00:16:44,551 --> 00:16:47,501
we can verify that our systems are working the way they're supposed to.

331
00:16:47,741 --> 00:16:49,991
You know, that's where I'm spending a lot of time thinking about right

332
00:16:49,991 --> 00:16:54,041
now is how do we use these tools that we were given to basically harden

333
00:16:54,042 --> 00:16:56,621
all of our production systems, both through the security stuff I was

334
00:16:56,621 --> 00:17:00,741
talking about earlier, as well as availability and, you know, testing.

335
00:17:01,042 --> 00:17:06,122
Well, well here's the question too, where I can take a look at any

336
00:17:06,162 --> 00:17:10,102
open source project now, or even any, any closed source product, and

337
00:17:10,112 --> 00:17:14,411
with enough time and poking of Claude Code and the tokens to back

338
00:17:14,412 --> 00:17:18,311
that up, it can spit out effectively a quote-unquote clean room build

339
00:17:18,481 --> 00:17:22,571
where, all right, I just rewrote your closed source thing as an open

340
00:17:22,571 --> 00:17:27,041
source implementation, or I have taken your open technically or source

341
00:17:27,041 --> 00:17:31,731
available technically code and now I have built a version that I

342
00:17:31,731 --> 00:17:35,901
can do whatever I want with because it is not a one-to-one copying.

343
00:17:35,912 --> 00:17:37,361
There's no code reuse here.

344
00:17:37,361 --> 00:17:40,031
It is a re-implementation from first principles.

345
00:17:40,391 --> 00:17:44,342
That has always been theoretically possible, but a massive amount of work.

346
00:17:44,431 --> 00:17:46,052
Now it's just a medium amount of tokens.

347
00:17:46,731 --> 00:17:47,892
How is that changing things?

348
00:17:48,689 --> 00:17:50,240
The differentiation that I still see, right?

349
00:17:50,240 --> 00:17:53,139
So you can make kind of the same case about Elastichash, the managed

350
00:17:53,139 --> 00:17:56,320
service I work on, and the differentiation that we really see is

351
00:17:56,330 --> 00:18:00,669
that we've been running the service at scale for well over a decade.

352
00:18:01,120 --> 00:18:03,730
And you don't get that by just pointing Claude at

353
00:18:03,909 --> 00:18:05,839
the API endpoints and say, "Hey, reproduce this.

354
00:18:05,850 --> 00:18:08,580
Reproduce what this is working on, how this, you know, behaves behind the

355
00:18:08,580 --> 00:18:11,290
scenes." Like, one of the things that we've been seeing inside Elastichash

356
00:18:11,290 --> 00:18:14,149
is we're trying to contribute more of our, you know, performance features,

357
00:18:14,149 --> 00:18:16,959
our efficiency features back into open source, 'cause we think that helps

358
00:18:16,959 --> 00:18:21,549
us make the overall system more stable, 'cause we get more eyes on the code.

359
00:18:21,750 --> 00:18:25,059
The value proposition of open source, that you having a collective group

360
00:18:25,059 --> 00:18:28,579
of people trying to make code better, is still true in the age of AI.

361
00:18:28,579 --> 00:18:31,420
Like, more people are reviewing it, more people are thinking about it,

362
00:18:31,420 --> 00:18:34,780
more people are trying to hypothesize and come up with improvements.

363
00:18:35,290 --> 00:18:39,249
So yes, the cost to write code has gone down a lot,

364
00:18:39,579 --> 00:18:42,429
but to be fair, my job has never been writing code.

365
00:18:42,429 --> 00:18:45,569
Like, I haven't been writing code for seven or eight years, right?

366
00:18:45,610 --> 00:18:48,189
Like I, back when I was, like, a college grad, I wrote

367
00:18:48,199 --> 00:18:51,379
a lot of code myself, and the fact that that's gone down

368
00:18:51,379 --> 00:18:54,079
significantly doesn't mean I'm, like, producing 10 times more.

369
00:18:54,090 --> 00:18:56,210
I like to make the joke that I can write code 10 times

370
00:18:56,210 --> 00:18:59,949
faster, but I'm about 20% more efficient, 'cause most of my

371
00:18:59,950 --> 00:19:02,470
job is just showing up to meetings and arguing with people.

372
00:19:02,829 --> 00:19:03,259
Me too.

373
00:19:03,259 --> 00:19:04,980
What ma- but the weird part is I'm not invited to

374
00:19:04,980 --> 00:19:06,619
those meetings, which is neither here nor there.

375
00:19:06,929 --> 00:19:08,209
Ah, but you're still appreciated.

376
00:19:08,209 --> 00:19:09,059
We're happy you're there.

377
00:19:09,090 --> 00:19:10,589
You bring some levity to it.

378
00:19:10,810 --> 00:19:11,679
Oh, exactly.

379
00:19:12,009 --> 00:19:12,939
I try.

380
00:19:12,990 --> 00:19:17,289
Um, I have seen a massive proliferation on, in various online

381
00:19:17,289 --> 00:19:20,919
fora of people vibe coding some SaaS thing, uh, where it's

382
00:19:20,919 --> 00:19:23,240
clear that they do not know the first thing about the deep

383
00:19:23,240 --> 00:19:27,010
scale problems of this space, and throwing it over the wall.

384
00:19:27,010 --> 00:19:29,989
And, and I wanna be clear here, I'm as guilty as anyone.

385
00:19:30,299 --> 00:19:33,549
In fact, if you go to deploybar.app, you can see something

386
00:19:33,549 --> 00:19:36,750
I built where my platypus hangs out in the macOS deploy

387
00:19:36,750 --> 00:19:40,689
bar and just has a persistent notification whenever GitHub,

388
00:19:40,689 --> 00:19:45,389
Vercel, or GitLab, uh, are doing a CI/CD run, and it's free.

389
00:19:45,649 --> 00:19:48,439
Now, if you wanna pay five bucks a month for it, Billy will stop

390
00:19:48,439 --> 00:19:51,759
making fun of you, or alternately, for the masochists out there,

391
00:19:51,889 --> 00:19:55,059
he'll really care and go much deeper into making fun of you.

392
00:19:55,239 --> 00:19:57,309
But the utility remains free.

393
00:19:57,310 --> 00:20:00,839
It's the snark and the cynicism, and I think the innovative business

394
00:20:00,840 --> 00:20:05,369
model of pay me or I won't be nice to you is, is kind of a good approach.

395
00:20:05,670 --> 00:20:09,569
But, but that's a bit of an edge case exception here, because

396
00:20:09,579 --> 00:20:12,530
so much of it is just I, I vibe coded this thing last night.

397
00:20:12,689 --> 00:20:14,169
Who knows if I'll maintain it or not?

398
00:20:14,329 --> 00:20:15,210
Pay me money, please.

399
00:20:15,997 --> 00:20:16,227
Yeah.

400
00:20:16,267 --> 00:20:18,818
I, I'm not very optimistic that those are gonna stick around.

401
00:20:18,868 --> 00:20:21,197
I mean, I'm sure some of them are, you know, finding unique

402
00:20:21,197 --> 00:20:24,287
markets and, you know, part of, you know, the whole startup world

403
00:20:24,287 --> 00:20:27,237
is trying to find product market fit while you still have cash.

404
00:20:27,618 --> 00:20:31,158
And so, you know, I, I do believe AI will help find…

405
00:20:31,158 --> 00:20:33,947
help companies find that product market fit faster.

406
00:20:34,277 --> 00:20:36,257
But I'm sure a lot of them are just gonna go nowhere, right?

407
00:20:36,257 --> 00:20:39,007
If your goal is to try to just be a, a shallow copy

408
00:20:39,007 --> 00:20:42,078
of AWS or GCP, it's gonna be very difficult, right?

409
00:20:42,127 --> 00:20:45,907
AWS has so much institutional knowledge about how this stuff works that it's

410
00:20:45,908 --> 00:20:49,137
gonna be, you know, 'cause the years we've run all this stuff in production,

411
00:20:49,137 --> 00:20:52,637
it's gonna be hard to try to copy it just by pointing Claude at it, right?

412
00:20:52,637 --> 00:20:55,287
Claude wasn't trained on, you know, a lot of this information,

413
00:20:55,317 --> 00:20:57,587
so it's trained on just kind of open source stuff.

414
00:20:57,598 --> 00:20:59,627
And a lot of that stuff is, you know, just

415
00:20:59,627 --> 00:21:01,238
random stuff that people wrote on GitHub.

416
00:21:01,418 --> 00:21:01,698
And

417
00:21:02,098 --> 00:21:04,937
I think that that's not terrible.

418
00:21:05,007 --> 00:21:08,507
I'm giving a talk somewhat soon, and a key thesis behind

419
00:21:08,507 --> 00:21:12,828
it is the idea that the AI bots are terrible at writing

420
00:21:12,867 --> 00:21:16,557
Terraform because there's no good Terraform out in the wild.

421
00:21:16,828 --> 00:21:19,087
Uh, I… To be clear, I've seen a lot of great Terraform,

422
00:21:19,098 --> 00:21:21,757
but it's always for companies that have learned what

423
00:21:21,857 --> 00:21:24,877
bad Terraform does and spent the time to explore it.

424
00:21:24,877 --> 00:21:27,937
I'm sorry, I should say OpenTofu now, but still, it's, it's the same

425
00:21:27,947 --> 00:21:33,327
principle where it, it struggles to do things correctly in that space.

426
00:21:33,668 --> 00:21:35,777
And when it comes to infrastructure at least, there's a blast

427
00:21:35,778 --> 00:21:39,047
radius here that there isn't as much in other disciplines.

428
00:21:39,737 --> 00:21:40,217
That's true.

429
00:21:40,248 --> 00:21:44,017
And I mean, the… I know the big trend these days is to build skills to

430
00:21:44,017 --> 00:21:47,397
help with that type of stuff, build MCCP servers that can provide these

431
00:21:47,397 --> 00:21:51,318
skills sort of on demand, and they help shape the LLM so that they do the

432
00:21:51,318 --> 00:21:54,327
right thing instead of relying too much on their, their training data.

433
00:21:54,927 --> 00:21:57,027
And, you know, we've seen some good adoption of that.

434
00:21:57,138 --> 00:22:00,577
Uh, for example, we have an engineer working on building a Valkey skill.

435
00:22:00,848 --> 00:22:03,858
So it's one of the problems that the LLMs have is knowing

436
00:22:03,858 --> 00:22:05,727
what's a Redis feature and what's a Valkey feature.

437
00:22:05,737 --> 00:22:08,887
It very quickly thinks they're… It like, you know, sometimes thinks

438
00:22:08,888 --> 00:22:12,367
Valkey features are in Redis and vice versa, uh, 'cause it doesn't…

439
00:22:12,807 --> 00:22:16,257
A lot of this information is in the foundational model, so it's in

440
00:22:16,257 --> 00:22:19,577
all the training data, um, and it struggles to differentiate them.

441
00:22:19,757 --> 00:22:22,428
And so skills help be like, "Hey, this is a Valkey feature. This is a

442
00:22:22,428 --> 00:22:26,828
Redis feature." And that stuff also would, you know, will hopefully solve

443
00:22:26,828 --> 00:22:29,648
a lot of these problems we have around stuff like Terraform and OpenTofu

444
00:22:30,894 --> 00:22:34,874
There is a future here where a lot of this stuff

445
00:22:34,874 --> 00:22:37,234
slips below the surface level of awareness.

446
00:22:37,244 --> 00:22:39,194
Now, that has been the case for a long time.

447
00:22:39,444 --> 00:22:44,413
Uh, we go back to the late '90s, and building and running a web server took

448
00:22:44,453 --> 00:22:49,104
an in-depth knowledge of GCC compiler flags and the better part of a week.

449
00:22:49,324 --> 00:22:52,594
Then RPM and Dpkg came out.

450
00:22:52,673 --> 00:22:55,503
Then Yum and Apt came out on top of them.

451
00:22:55,703 --> 00:22:59,413
Then things like Puppet and Chef and whatnot came out, and it was simply

452
00:22:59,413 --> 00:23:03,063
just, you know, ensure installed, and then it became a checkbox on S3.

453
00:23:03,453 --> 00:23:06,883
And so things that were hard today become easier yesterday.

454
00:23:06,894 --> 00:23:08,563
That has been an ongoing trend.

455
00:23:08,874 --> 00:23:11,564
I guess I didn't think that I would necessarily live to

456
00:23:11,564 --> 00:23:15,773
see a world where an entire app fit into that bucket.

457
00:23:16,833 --> 00:23:19,593
I don't know how much you were ever using, like, the early GPT models,

458
00:23:19,594 --> 00:23:23,694
but I used to use GPT-2 and GPT, uh, the early versions of GPT-3

459
00:23:23,743 --> 00:23:28,294
that OpenAI produced, uh, to help do, like, D&D tabletop campaigns,

460
00:23:28,313 --> 00:23:32,773
and, like, that's about a decade ago when this came out, and the rate

461
00:23:32,773 --> 00:23:36,413
at which they've kind of been evolving is faster than I expected.

462
00:23:36,773 --> 00:23:39,304
But I imagine we'll continue seeing that type of

463
00:23:39,304 --> 00:23:41,734
progress, and yeah, as you said, I think those types of

464
00:23:41,763 --> 00:23:44,633
workloads will become commodities in not too much time.

465
00:23:45,673 --> 00:23:46,664
What do you think is next?

466
00:23:46,673 --> 00:23:51,434
Because I, and I wanna be clear here, five years ago, if you had accurately

467
00:23:51,453 --> 00:23:56,123
predicted what the current state of the world is in terms of open source,

468
00:23:56,123 --> 00:23:59,853
in terms of software development, you would have sounded like a lunatic.

469
00:24:00,123 --> 00:24:02,754
With that in mind, what do you think the next five years looks like?

470
00:24:03,283 --> 00:24:06,663
I do hate making predictions, but I'll, I'll do my best.

471
00:24:06,853 --> 00:24:10,084
I definitely see an increasing trend of the, you know, like, the

472
00:24:10,084 --> 00:24:13,074
cost of writing code continues to come down, and by cost, I mean

473
00:24:13,074 --> 00:24:15,653
both, you know, in terms of token, how much time engineers are being

474
00:24:15,653 --> 00:24:19,623
spent, but, like, the high-context individuals who are driving a lot

475
00:24:19,623 --> 00:24:22,973
of this stuff still are driving it like they were a few years ago.

476
00:24:23,493 --> 00:24:28,934
So I kind of do expect to see just more and more, like,

477
00:24:29,233 --> 00:24:33,758
high-velocity Individuals sort of making a lot of stuff happen.

478
00:24:34,168 --> 00:24:36,418
So like the people that are able to, you know… Like

479
00:24:36,447 --> 00:24:39,328
the Alaska service I imagine will have fewer people, but

480
00:24:39,328 --> 00:24:41,717
they're able to drive more stuff, build more features.

481
00:24:42,407 --> 00:24:44,687
And I imagine more of that will happen both

482
00:24:44,687 --> 00:24:46,518
like in the startup world and individuals.

483
00:24:46,748 --> 00:24:50,457
I haven't used like OpenClaw or those types of tools too much for

484
00:24:50,457 --> 00:24:53,058
my personal life, but I imagine that will become more ubiquitous.

485
00:24:53,578 --> 00:24:58,847
Like, I think we'll see a lot of just the same, of just, you know, being

486
00:24:58,848 --> 00:25:02,267
able to force multiply through AI, having it be easier to do, have them

487
00:25:02,268 --> 00:25:05,017
all… Like, one of the big innovations that happened recently was just

488
00:25:05,017 --> 00:25:10,327
like you could basically prompt the Claude code and it will figure it out.

489
00:25:10,357 --> 00:25:13,198
I imagine that'll continue to happen sort of in all aspects of life.

490
00:25:13,537 --> 00:25:16,428
But like that's kind of what I see happening kind of across the next five years.

491
00:25:17,278 --> 00:25:19,548
Like this is the first time I think I would ever say that I

492
00:25:19,557 --> 00:25:21,417
don't really know what's gonna happen five years from now.

493
00:25:21,807 --> 00:25:22,037
Yeah.

494
00:25:22,067 --> 00:25:24,927
I'm, I'm hoping it sorts itself out before the time my elementary

495
00:25:24,927 --> 00:25:28,358
school kids have to enter the workforce, but we'll find out.

496
00:25:28,528 --> 00:25:31,987
Uh, I, I will say you can probably make some better predictions because,

497
00:25:32,028 --> 00:25:36,487
again, Valkey is open source and it is not AWS controlled or restrained.

498
00:25:36,967 --> 00:25:37,958
What's coming in Valkey?

499
00:25:38,517 --> 00:25:40,887
That is much easier to make predictions about.

500
00:25:41,117 --> 00:25:43,007
So the Valkey project, the main things we're

501
00:25:43,007 --> 00:25:46,388
working on are basically improving durability.

502
00:25:46,397 --> 00:25:49,568
So Valkey has, Valkey and Redis have historically had a durable

503
00:25:49,577 --> 00:25:52,907
version called append-only files, which was pretty much self-instance.

504
00:25:53,197 --> 00:25:55,898
We're trying to make it a durable multi-node distributed

505
00:25:55,898 --> 00:25:58,737
system, and this will allow people to actually run stuff.

506
00:25:58,737 --> 00:26:01,168
Like we kind of want to replace Kafka workloads.

507
00:26:01,727 --> 00:26:03,577
We would love to replace stuff like some very

508
00:26:03,577 --> 00:26:06,197
simple key value primary data store workloads.

509
00:26:06,327 --> 00:26:08,977
There's some interesting use cases and stuff like vector similarity

510
00:26:08,977 --> 00:26:11,487
search where you actually do want the indexes durably committed.

511
00:26:11,647 --> 00:26:13,467
So those are the type of things that we're building out.

512
00:26:13,787 --> 00:26:15,977
We're trying to also… You know, the DRAM

513
00:26:15,977 --> 00:26:18,467
shortage is, you know, on everybody's mind.

514
00:26:18,537 --> 00:26:21,088
One of the big things that's been on the Valkey project's

515
00:26:21,107 --> 00:26:23,807
roadmap for a long time is figuring out how to natively store

516
00:26:23,837 --> 00:26:27,397
data onto SSDs without impacting latency and performance.

517
00:26:27,857 --> 00:26:31,177
There's been a lot of innovations in the last couple of years that have made

518
00:26:31,227 --> 00:26:37,383
SSD read latencies like Very competitive with RAM, like sub 10 mil- microsecond

519
00:26:37,393 --> 00:26:40,974
reads, um, which is still 100 times slower than DRAM, but as long as you

520
00:26:40,974 --> 00:26:44,954
carefully orchestrate how you're fetching the data, it's can almost be free.

521
00:26:45,373 --> 00:26:47,444
So those are the two big things we're working on as a project.

522
00:26:48,063 --> 00:26:50,553
But yeah, we're al- always hopeful to get more things.

523
00:26:50,563 --> 00:26:51,643
Those are features that are probably gonna

524
00:26:51,643 --> 00:26:53,093
come out in the next six to 12 months.

525
00:26:53,493 --> 00:26:54,664
And of course, there's a lot of other stuff.

526
00:26:54,664 --> 00:26:56,093
We're releasing Valkey 9.1.

527
00:26:56,113 --> 00:26:58,633
It should be out by the time this pod comes, podcast comes out.

528
00:26:58,633 --> 00:27:01,423
That adds performance improvements, memory efficiency stuff.

529
00:27:01,444 --> 00:27:02,933
Those are the bread and butter that-

530
00:27:02,934 --> 00:27:04,263
Well, that's putting an awful lot of faith in

531
00:27:04,283 --> 00:27:06,623
GitHub, uh, staying up long enough to ship a release.

532
00:27:06,693 --> 00:27:07,683
I'm sorry, that's unkind.

533
00:27:07,853 --> 00:27:10,423
Fair, 'cause it's very expensive, but unkind.

534
00:27:10,423 --> 00:27:12,293
I, I sh- I wasn't gonna complain about GitHub,

535
00:27:12,303 --> 00:27:14,214
but I'm still not gonna complain about GitHub.

536
00:27:14,233 --> 00:27:16,154
Uh, all the current problems we're having are our own problems.

537
00:27:16,154 --> 00:27:19,024
We're trying to release when… As, as of recording of this podcast, we're

538
00:27:19,024 --> 00:27:23,194
trying to do a, a, a patch release of Valkey and it is taking forever.

539
00:27:23,363 --> 00:27:23,843
I'll be direct.

540
00:27:23,853 --> 00:27:26,454
My problem with GitHub right now is they are simultaneously

541
00:27:26,454 --> 00:27:29,033
saying that it is not their fault because they are being slammed

542
00:27:29,033 --> 00:27:32,073
by a deluge of AI stuff, which I believe and it's sincere.

543
00:27:32,303 --> 00:27:35,163
However, they're also shoving Copilot at anything that

544
00:27:35,163 --> 00:27:37,073
holds still long enough, and many things that don't.

545
00:27:37,303 --> 00:27:40,293
So it's, it feels like you, you don't get to

546
00:27:40,354 --> 00:27:43,123
sell the problem and then complain about it.

547
00:27:43,463 --> 00:27:44,914
I mean, that's capitalism, right?

548
00:27:45,123 --> 00:27:46,263
Oh, it is.

549
00:27:46,534 --> 00:27:50,194
I just… It's, I, I sit here and I shake my fist and it makes me angry.

550
00:27:50,854 --> 00:27:54,123
If, if people wanna learn more about what you're up to and what's next in the

551
00:27:54,123 --> 00:27:57,064
exciting world of Valkey, where's the best place for them to go to find you?

552
00:27:57,253 --> 00:27:59,493
Best place to find me is probably on Bluesky.

553
00:27:59,493 --> 00:28:02,393
I'm reconditerose, uh, bluesky.social.

554
00:28:02,603 --> 00:28:05,163
You following the Valkey project, valkey.io, there's

555
00:28:05,183 --> 00:28:08,793
a blog section which we publish relatively frequently.

556
00:28:08,793 --> 00:28:12,703
We've almost gotten enough content that we have a weekly blog release

557
00:28:12,704 --> 00:28:15,793
cadence about everything new and exciting on the Valkey project.

558
00:28:16,283 --> 00:28:18,954
Uh, LinkedIn is also a great place to follow both Valkey and me.

559
00:28:19,013 --> 00:28:21,984
I mostly just repost stuff, but it's kind of what's going on in the project.

560
00:28:22,343 --> 00:28:24,753
We're, we're planning to ho- uh, host a Valkey

561
00:28:24,793 --> 00:28:27,554
event, uh, at the Open Source Summit in May.

562
00:28:27,603 --> 00:28:31,403
That will probably not… Maybe we'll be in time, but more

563
00:28:31,403 --> 00:28:34,833
likely we're hosting a, an event called Unlocked this week, but

564
00:28:34,833 --> 00:28:37,383
there's also gonna be another one in Prague, hopefully in Q3.

565
00:28:37,753 --> 00:28:38,833
So we're trying to organize that.

566
00:28:38,833 --> 00:28:41,093
So if you're interested in coming and learning a lot more about

567
00:28:41,093 --> 00:28:44,653
Valkey, we're planning on hosting an event there later this year.

568
00:28:45,443 --> 00:28:48,993
Wonderful, and we will of course put links to that in the show notes.

569
00:28:49,384 --> 00:28:52,143
Madeline, thank you so much for taking the time to speak with me.

570
00:28:52,193 --> 00:28:55,354
As always, it is a pleasure and I'm looking forward to the next time.

571
00:28:55,943 --> 00:28:56,493
Excellent.

572
00:28:56,513 --> 00:28:58,633
Hope I, hope I wasn't too AI-pilled for you.

573
00:29:00,143 --> 00:29:00,653
Not yet.

574
00:29:00,663 --> 00:29:03,654
That's okay though, 'cause I'm gonna write conspiracy theories about it myself.

575
00:29:04,363 --> 00:29:09,293
Madeline Olson, AWS Principal Engineer and core maintainer of Valkey.

576
00:29:09,634 --> 00:29:13,673
I'm cloud economist Corey Quinn, and this is Screaming in the Cloud.

577
00:29:13,943 --> 00:29:15,804
If you've enjoyed this podcast, please leave a

578
00:29:15,804 --> 00:29:18,223
five-star review on your podcast platform of choice.

579
00:29:18,483 --> 00:29:21,554
Whereas if you hated this podcast, please leave a five-star

580
00:29:21,554 --> 00:29:24,503
review on your podcast platform of choice, along with

581
00:29:24,503 --> 00:29:28,253
an angry comment that no doubt will present as AI slop.