Debugging simple memory leaks in go #sre
栏目分类：资料 发布日期：2018-08-03 浏览次数：次
本文为去找网小编(www.7zhao.net)为您推荐的Debugging simple memory leaks in go #sre，希望对您有所帮助，谢谢！
are a class of bugs where memory is not released even after it is no longer needed. They are often explicit, and highly visible, which makes them a great candidate to begin learning debugging. Go is a language particularly well suited to identifying memory leaks because of its powerful toolchain, which ships with amazingly capable tools ( ) which make pinpointing memory usage easy.
I’m hoping this post will illustrate how to visually identify memory, narrow it down to a specific process, correlate the process leak with work, and finally find the source of the memory leak using pprof. This post is intentionally contrived in order to allow for the simple identification of the root cause of the memory leak. The pprof overview is intentionally brief and aims to illustrate what it is capable of and isn’t an exhaustive overview of its features. 欢迎访问www.7zhao.net
What is a memory leak?
If memory grows unbounded and never reaches a steady state then there is probably a leak. The key here is that memory grows without ever reaching a steady state, and eventually causes problems through explicit crashes or by impacting system performance. 内容来自www.7zhao.net
Memory leaks can happen for any number of reasons. There can be logical leaks where data-structures grow unbounded, leaks from complexities of poor object reference handling, or just for any number of other reasons. Regardless of the source, many memory leaks elicit a visually noticeable pattern: the “ sawtooth ”.
This blog post is focused on exploring how to identify and pinpoint root cause for a go memory leak. We’ll be focusing primarily on characteristics of memory leaks, how to identify them, and how to determine their root cause using go. Because of this our actual debug process will be relatively superficial and informal. 欢迎访问www.7zhao.net
The goal of our analysis is to progressively narrow scope of the problem by widdling away possibilities until we have enough information to form and propose a hypothesis.After we have enough data and reasonable scope of the cause, we should form a hypothesis and try to invalidate it with data.
Each step will try to either pinpoint a cause of an issue or invalidate a non-cause. Along the way we’ll be forming a series of hypotheses, they will be necessarily general at first then progressively more specific. This is loosely based on the scientific method. Brandon Gregg does an amazing job of covering different for system investigation (primarily focused on performance).
Just to reiterate we’ll try to:
- Ask a question
- Form a Hypothesis
- Analyze the hypothesis
- Repeat until the root cause is found
How do we even know if there is a problem (ie memory leak)? Explicit errors are direct indicators of an issue. For memory leaks a common error are OOM errors or explicit system crashes.
Errors are the most explicit indicator of a problem. User generated errors have the potential to generate false positives if the logic is off OOM is the OS literally indicating something is using too much memory. In the error below this manifests as cgroup limits being reached and the container being killed. 本文来自去找www.7zhao.net
Question:Is the error a regular repeating issue?
Hypothesis:OOM errors are significant enough that they should rarely occur. There is a memory leak in one of the processes. www.7zhao.net
Prediction:Either the Process memory limit has been set too low and there was a uncharacteristic bump or there is a larger issue.
Test:Upon further inspection there are quite a few OOM errors suggesting this is a serious issue and not a one off. Check the system memory for historic view into memory usage. www.7zhao.net
The next step after identifying a potential problem is to get an idea of system wide memory usage. Memory Leaks frequently display a “sawtooth” pattern. The spikes correspond to the application running while the dips correspond to a service restart. 欢迎访问www.7zhao.net
Sawtooth characterizes a memory leak especially corresponding with a service deploy. I’m using a test project to illustrate memory leaks but even a slow leak would look like saw tooth if the range is zoomed out far enough. With a smaller time range it would look like a gradual rise and then a drop off on process restart. www.7zhao.net
The graph above shows the example of a sawtooth memory growth. Memory continually grows without flatlining. This is a smoking gun for memory issues.
Question:Which process (or processes) is (are) responsible for the memory growth?
Test:Analyze per process memory. There could be information in the dmesg logs to indicate a process or class of processes that are the targets of OOM. 本文来自去找www.7zhao.net
Per Process Memory
Once a memory leak is suspected the next step is to identify a process that is contributing, or causing the system memory growth. Having per process historical memory metrics is a crucial requirement (container based system resources are available through the a tool like ). Go’s provides per process memory metrics by default, which is where the graph below gets its data. copyright www.7zhao.net
The below graph shows a process which is very similar to the system sawtooth memory leak graph above: continual growth until process restarts.
Memory is a critical resource and can be used to indicate abnormal resource usage or could be used as a dimension for scaling. Additionally, having memory stats help inform how to set container based ( ) memory limits. The s of the graph values above can be found . After the process has been identified it’s time to dig in and find out which specific part of the code is responsible for this memory growth.
Root Cause Analysis / Source Analysis
Go Memory Analysis
Once again prometheus gives us detailed information into the go runtime, and what our process is doing. The chart shows that bytes are continually allocated to the heap until a restart. Each dip corresponds with the service process restart. copyright www.7zhao.net
Question:Which part(s) of the application is(are) leaking memory?
Hypothesis:There’s a memory leak in a routine which is continually allocating memory to the heap (global variable or pointer, potentially visible through )
Test:Correlate the memory usage with an event. 本文来自去找www.7zhao.net
Correlation With Work
Establishing a correlation will help to partition the problem space by answering: is this happening online (in relation to transactions) or in the background?
One way to determine this could be to start the service and let it idle without applying any transactional load. Is the service leaking? If so it could be the framework or a shared library. Our example happens to have a strong correlation with transactional workload.
The above graph show the count of HTTP requests. These directly match the system memory growth and time and establish diving into HTTP request handling as a good place to start.
Question:Which part of the application are responsible for the heap allocations?
Hypothesis: There is an HTTP handler that is continually . www.7zhao.net
Test:Periodically analyze heap allocations during program running in order to track memory growth. www.7zhao.net
Go Memory Allocations
In order to inspect how much memory is being allocated and the source of those allocations we’ll use . pprof is an absolutely amazing tool and one of the main reasons that I personally use go. In order to use it we’ll have to first enable it, and then take some snapshots. If you’re already using http, enabling it is literally as easy as: www.7zhao.net
import _ "net/http/pprof" 内容来自www.7zhao.net
Once pprof is enabled we’ll periodically take heap snapshots throughout the life of process memory growth. Taking a heap snapshot is just as trivial:
curl http://localhost:8080/debug/pprof/heap > heap.0.pprof sleep 30 curl http://localhost:8080/debug/pprof/heap > heap.1.pprof sleep 30 curl http://localhost:8080/debug/pprof/heap > heap.2.pprof sleep 30 curl http://localhost:8080/debug/pprof/heap > heap.3.pprof 本文来自去找www.7zhao.net
The goal is to get an idea of how memory is growing throughout the life of the program. Let’s inspect the most recent heap snapshot: 欢迎访问www.7zhao.net
This is absolutely amazing. pprof defaults to
Type: inuse_space which displays all the objects that are currently in memory at the time of the snapshot. We can see here that
bytes.Repeat is directly responsible for 98.60% of all of our memory!!!
The line below that shows:
1.28MB 0.31% 98.91% 410.25MB 98.91% main.(*RequestTracker).Track
This is really interesting, it shows that
Track itself has
0.31% but is responsible for
98.91% of all in use memory!!!!!!!!!!!!! Further more we can see that http has even less memory in use but is responsible for even more than
Track is called from it).
pprof exposes many ways to introspect and visualize memory (in use memory size, in use number of objects, allocated memory size, allocated memory objects), it allows listing the track method and showing how much each line is responsible for:
This directly pinpoints the culprit:
1.28MB 410.25MB 24: rt.requests = append(rt.requests, bytes.Repeat(byte("a"), 10000))
pprof also allows visual generation of the textual information above:
(pprof) svg Generating report in profile003.svg内容来自www.7zhao.net
This clearly shows the current objects occupying the process memory. Now that we have the culprit
Track we can , and fix the root issue.
Resolution: Memory was being continually allocated to a global variable on each HTTP request. The memory growth in this case should have been inspected. www.7zhao.net
I hope that this post was able to illustrate the power of visually identifying memory leaks and a loose process for progressively narrowing down its source. Finally, I hope that it was able to touch on the power of pprof for go memory introspection and analysis. As always I would love any feedback. Thank you for reading. copyright www.7zhao.net
以上为Debugging simple memory leaks in go #sre文章的全部内容，若您也有好的文章，欢迎与我们分享！ 去找(www.7zhao.net欢迎您
- iOS 相机流人脸识别(二)-关键点检测(face l
- Bridging the gap between low-level and high-level Jav
- Going cloud-native costs more than you think
- Reading Waterfall Charts to Focus on Page Speed
- Why Only Smart People Buy Your Products
- HugeGraph – An open source fast and highly scalable
- iOS端近场围栏检测 ——MultipeerConnectivity