Datasets:
willtheorangeguy
/
2021-Ship-It-Transcripts

Dataset card Data Studio Files
xet
 Community
text
stringlengths
0
1.8k
[1708.78 --> 1710.44] There had been kind of two problems with it.
[1710.44 --> 1716.28] One was, for the longest time, profiling was incredibly expensive to do in production.
[1716.28 --> 1725.16] You would only do it to specific processes at a certain, like, on-demand because you didn't want to create too much additional overhead.
[1725.60 --> 1731.58] There was one thing that kind of led to us being able to do this in production and always on.
[1732.00 --> 1735.68] And one of those things is what we call sampling profiling.
[1735.68 --> 1748.90] So, instead of kind of tracing exactly, absolutely everything a process does, we only, you know, 100 times per second look at what the program does at that particular moment in time and capture the stack trace of what it does.
[1749.12 --> 1752.74] Because, essentially, the stack trace represents what the program is doing, right?
[1752.74 --> 1764.02] And so, for a start already, for some hyperscalers, this was already enough to build continuous profiling tools for them to consume internally because they could do it always on in production now.
[1764.34 --> 1771.42] Now, as it goes with so much cloud-native technology and developments, that wasn't necessarily accessible to everyone.
[1771.42 --> 1778.44] And one of the really amazing things that also have happened somewhat recently has been eBPF.
[1778.82 --> 1790.00] And eBPF allows us to capture this data at an even lower overhead because we can already capture it in the form that we are going to consume it afterwards.
[1790.42 --> 1798.70] We don't need to use some pre-baked format that may have a ton of information that we don't need, a ton of detail we don't need.
[1798.70 --> 1805.96] We can produce exactly the data that we want and make that exportable to user space and then ingest it into our storage.
[1806.28 --> 1810.56] So, that was definitely also a really big part of what created a movement.
[1811.16 --> 1813.78] But kind of, this doesn't really have to do with overhead.
[1814.38 --> 1821.80] There's also another aspect, which is just kind of Kubernetes unifying the observability space in a way.
[1821.80 --> 1825.82] And I think we might have talked about this in our last session, actually.
[1826.36 --> 1832.08] The way that Prometheus also and Kubernetes have kind of standardized a lot of terms in our industry.
[1832.36 --> 1835.24] It just makes us all speak the same language.
[1835.86 --> 1843.78] And so, this is super powerful because all of a sudden, when I say pod and you say pod, we immediately know what we're talking about, right?
[1843.78 --> 1850.62] And so, this is much more cultural than it is technologically, but it means that our knowledge is transferable.
[1851.04 --> 1852.46] And so, this is incredibly powerful.
[1852.70 --> 1855.40] And then the last piece is putting all of this together.
[1855.94 --> 1862.50] eBPF with Kubernetes now allows us to automatically discover all of the containers that are running in our infrastructure
[1862.50 --> 1868.88] and be able to look at all the CPU time that is being consumed in our infrastructure at once.
[1868.88 --> 1880.66] And the reason why this is so powerful is because all of a sudden, we can now say this stack trace in this binary is what's causing 20% of our CPU time.
[1881.14 --> 1887.76] If we optimize this stack trace away, we're now saving 20% of CPU time in our infrastructure.
[1888.26 --> 1889.54] That's huge, right?
[1889.60 --> 1895.16] Think of the banks, automotive companies, any company that has a large cloud bill, right?
[1895.48 --> 1898.62] They can save millions of dollars with these kinds of measurements.
[1898.62 --> 1902.10] It's just the reality is they can do these measurements today.
[1902.44 --> 1905.86] And it doesn't really matter what language you're using, right?
[1905.90 --> 1907.94] Because everything runs as a pod.
[1908.20 --> 1911.18] It doesn't matter whether it's Java, whether it's Go, whether it's Erlang.
[1911.30 --> 1912.14] It really doesn't matter.
[1912.56 --> 1920.04] The point being is you run this agent on your Kubernetes worker node where all these pods are being scheduled.
[1920.46 --> 1924.96] And you can see out of the pods which are being scheduled, out of the containers which are running within those pods,
[1925.30 --> 1927.42] which are the ones that consume the most CPU.
[1927.42 --> 1929.48] And I imagine this goes beyond CPU.
[1929.96 --> 1939.14] It goes to memory, disk operations, network operations, I operations, all that nice, important stuff that the kernel knows about.
[1939.30 --> 1942.76] And it presents you via eBPF in a way that makes sense to you.
[1942.84 --> 1945.94] And it doesn't matter what language is making that call.
[1946.10 --> 1948.50] Whether you have a serverless framework, it really doesn't matter.
[1948.60 --> 1949.44] It's really powerful.
[1949.74 --> 1951.18] I like the way you're thinking about this.
[1951.18 --> 1958.14] So I was going to ask you, parka.dev is the thing that you're opening up to the world at this KubeCon.
[1958.50 --> 1962.54] And I was going to ask you, why do you need parka?
[1962.74 --> 1964.90] But I think the answer is to cost optimize.
[1965.62 --> 1967.02] But maybe there's something more to it.
[1967.24 --> 1970.30] First of all, I think, and we said this in our announcement as well.
[1970.30 --> 1979.52] I think just the people that we are and the company that we're building, I think we needed to have an open source piece to be ourselves.
[1979.52 --> 1986.56] So even if there wasn't anything else, that would probably already would have been enough of an argument for us.
[1986.94 --> 1997.72] But I think more importantly, continuous profiling is, even though there are now several vendors, several projects out there, in the only one year that Polar Signals has existed, right?
[1997.80 --> 2002.34] Like there are several companies that have sprung up, several vendors that have created products.
[2002.66 --> 2007.02] But it's still a really young space and is still not very well understood.
[2007.02 --> 2017.26] And so in a way, the open source project is also about democratizing this for the community and educating the community about continuous profiling.
[2017.56 --> 2024.50] So that when we talk about continuous profiling, hopefully in a year or two, everyone understands it like when I say distributed tracing.
[2024.90 --> 2031.12] So if I understand correctly, it's your need to understand what the system does.
[2031.12 --> 2037.12] And the itch that you're scratching is you wanting to understand what is happening on those nodes.
[2037.52 --> 2038.40] So that's why I did it.
[2038.68 --> 2039.30] As simple as that.
[2040.18 --> 2040.80] I love that.
[2041.20 --> 2041.74] I love that.
[2041.88 --> 2045.72] The backstory actually goes a little bit further than where I started.
[2045.72 --> 2057.14] But the reason why I even went into putting together that proof of concept with Conprof was because I read a paper by Google where they described these methodologies, right?
[2057.14 --> 2064.02] How they used these kind of methods to cut down on infrastructure costs every quarter by multiple percentage points, right?
[2064.46 --> 2067.64] And I was just amazed by that for several reasons.
[2068.00 --> 2072.38] One, I just wanted to have this tool while I was working on Prometheus, right?
[2072.38 --> 2075.62] And the other one was I had worked on Prometheus.
[2076.02 --> 2082.20] At least I thought to myself, I think I know a thing or two about working with data over time, right?
[2082.26 --> 2089.64] And so I think that's kind of what ultimately created the circumstances of me wanting to create a tool like this.
[2089.98 --> 2092.82] So I got the tool up and running in seconds.
[2092.96 --> 2093.08] Right.
[2093.22 --> 2095.78] Like that just shows how easy it is to get started.
[2095.86 --> 2096.76] This was just local.
[2097.16 --> 2101.90] I didn't want to venture in our production Kubernetes cluster because I have something else in mind for that.
[2101.90 --> 2104.94] But in a few seconds, I could access the UI.
[2105.28 --> 2106.52] I could see the CPU time.
[2107.12 --> 2113.36] And the UI, what surprised me, is it's better than the first Prometheus UI that I remember.
[2113.72 --> 2117.24] And I think the secret to this is your coffee machine.
[2117.46 --> 2118.12] Let me explain.
[2118.32 --> 2118.54] Okay.
[2118.60 --> 2119.18] Let me explain.
[2119.84 --> 2121.46] So this is what's going on in my mind.
[2122.10 --> 2130.34] When I first heard of Parca a few weeks back, I checked it out and it was looking good, but it wasn't as polished as it is today.
[2130.34 --> 2136.26] Just in a matter of a few weeks, I was astounded by how fast you're iterating on it.
[2136.64 --> 2139.06] And I think that it's your new coffee machine.
[2140.06 --> 2141.28] Is that it?
[2141.82 --> 2142.68] What's the secret?
[2143.04 --> 2144.76] I would say it has a part in it.
[2145.10 --> 2145.42] Okay.
[2145.42 --> 2151.00] I think the UI is actually an evolution of several attempts at it.
[2151.22 --> 2171.78] The very first one was actually within our closed source beta product where, you know, when we launched it in early February this year, we used this to work really closely with a couple of early users to understand what is it that they, beyond the UI even, what is it that they want from an experi...
[2171.78 --> 2172.06] Right.
[2172.44 --> 2178.36] But then also, of course, like also with ourselves using the software, like how do we want to use it?
[2178.46 --> 2186.40] And so I think there's so much dogfooding that was going on from basically day one, because this is a tool that we built for ourselves.
[2186.64 --> 2188.66] We wanted to put that work into it.
[2188.82 --> 2188.92] Right.
[2189.18 --> 2190.22] What do you use the tool for?
[2190.28 --> 2190.98] This is really interesting.