Hi Paul,
My comments in-line.
Thank you,
Vlad
On 4/5/18 20:50, Paul Rogers wrote:
Hi Vlad,
I'd suggest to keep focus on DrillBuf design and implementation as the only
gate for accessing raw (direct) memory.
I was doing that. By explaining where DrillBuf fits in the overall design, we
see that DrillBuf should be the only access point for direct memory. The
context explains why this is the right decision. Changes to DrillBuf should
support our design as DrillBuf only exists for that purpose.
My concern is not why Drill adopted Netty as Netty does provide a good
amount of functionality on top of Java networking and NIO. I also do not
propose to replace Netty with something else. My primary focus for this
thread is the design and implementation of the DrillBuf Java class
itself. Namely, why was it necessary to introduce DrillBuf and/or
UnsafeDirectLittleEndian? What functionality do those classes provide
that existing Netty classes do not? Netty already provides memory
pooling, reference counting, slicing, composite buffers, working with
direct and heap memory. By looking at the DrillBuf.java git history, the
DrillBuf was introduced in 2014 and prior to that Netty classes were
used directly. Unfortunately, the commit that introduced DrillBuf does
not provide any info why it was introduced and does not have a reference
to a JIRA.
One may argue that DrillBuf is a way for Drill to encapsulate Netty
ByteBuf and guard other modules that use DrillBuf from Netty ByteBuf
API, so if Netty decides to change ByteBuf API in the next major
release, amount of changes will be limited to DrillBuf only. Problem is
that DrillBuf inherits from Netty AbstractByteBuf, so the above goal is
not achieved either.
1. Boundary checking (on/off based on a flag or assertions being
enabled/disabled, always on, always off, any other suggestions)
By understanding the design, we can see that we do, in fact, need both checked
and unchecked methods. The row set mechanisms takes it upon themselves to have
sufficient context to ensure that memory access is always within bounds, and so
can benefit from the unchecked methods. As we said, we need debug-time bounds
checks to catch errors during development.
On the other hand, value vectors should probably be protected by using checked
methods because they do not have intrinsic mechanisms that ensure correct
access. With vectors, the memory location to access is set by the caller (each
operator) and there is no guarantee that all this code is correct all the time.
(Though, it probably is right now because if it wasn't we'd get errors.)
I don't see a difference in bounds checking requirements between row set
mechanism and value vectors as value vectors do have "safe" methods or
intrinsic mechanism that ensures correct access. If not all operators
use "safe" methods, than that operator should provide a guarantee. At
the end, if an operator accesses memory out of allocated bounds,
boundary checking will not fix it. If there is a bug in row set
mechanism, value vectors or an operator, the end result (crash of the
JVM) is the same for all.
The proposal just made represents a change; currently the two mechanisms use the same set
of methods, which puts us into the "should we turn on bounds checks for everyone or
turn them off for everyone" dilemma.
This is a technical design decision, not a community preference (other than
that we'd prefer that stuff works...)
On the dev list, almost everything is a technical design decision and
the community (compared to Drill users who prefer that the community
makes a choice and stuff works) needs to agree on how to proceed and
what development practices to adopt and follow.
2. Ref count checking (delegate to netty or have a separate mechanism to
enable/disable, always on or off)
Ref counts are absolutely necessary, in the current design, for the reasons
explained earlier: a single memory block can be shared by multiple DrillBufs.
We have no other way at present to know when the last reference goes away if we
don't have ref counts.
To deprecate reference counts, we'd have to rework the way that memory is
transferred between operators. We'd have to deprecate shared buffers. (Or, we'd
have to move to the fixed blocks mentioned earlier; but even then we'd need ref
counts if a single sender can feed data to multiple fragments without copies.)
Again, this is not a preference issue, it is a fundamental design issue (unless
you know of a trick to remove the need for ref counts, in which case please do
propose it.) Or, if there is a better way to implement bounds checks that is
faster or simpler, please do propose that.
The question is not about deprecating reference count mechanism.
Question is whether or not to check that reference count is not zero
every time DrillBuf is used (see ensureAccessible()).
3. Usage of UDLE
If we meet the design goals stated earlier, and DrillBuf is the only interface
to memory, then we could change the internal representation if there were a
good reason to do so. Since you listed UDLE, you probably have an idea in mind.
To make the code simpler? Faster? Perhaps make a proposal.
IMO, it will be good to make DrillBuf code simpler and consistent. As
the first step, it will be good to understand existing design and
implementation in the absence of comments and/or JIRA. For example,
where UnsafeDirectLittleEndian.getLong(int index) is used? Or, why
UnsafeDirectLittleEndian overrides "public int setBytes(int index,
InputStream in, int length)" and does not override "public ByteBuf
setBytes(int index, ByteBuf src, int length)"?
4. Changing DrillBuf to follow Netty convention
Why? Vectors are not byte buffers; they just happen to use them as explained
earlier. Vectors are actually, well, vectors of fixed types with rigid
structure, so DrillBuf should provide methods that implement that design.
Further, Netty never uses DrillBuf so who would benefit from an API change?
Just to make the code cleaner? To gain performance? Something else? Again,
proposal? Note that the comment about checked and unchecked methods made
earlier suggests we'd make DrillBuf even *less* like Netty...
I am not sure I understand your reference to Vectors, can you please
explain. I do not propose to change DrillBuf to follow Vectors API. The
proposal is to make DrillBuf methods that override Netty methods to be
consistent with the rest of Netty API. Note that it is not correct that
Netty never uses DrillBuf API. It does when DrillBuf delegates to
UnsafeDirectLittleEndian or does not override inherited methods. In many
cases Netty implementation relies on abstract methods implemented in
subclasses. Netty also has unchecked variants and checked methods use
unchecked most of the time.
5. Moving DrillBuf to a different package
This is purely a functional question. As the lore has it, DrillBuf is in the
Netty package because it needs (or needed) visibility to protected members in
Netty classes. If that is no longer true, then lets do move it to the Drill
package so we confuse ourselves less about what is and what is not Drill code.
(I'm STILL confused...) Have you checked to see if things compile if we do the
move?
There is no need to place a class into the same package to use protected
members of the base class. There may be classes or fields with package
private visibility and IMO, it is better not to use them (see my
previous e-mail). Again, I do not want to surprise the community with
JIRA and/or PR with massive package name changes and I would like to get
the community consensus before implementing those changes.
Thanks,
- Paul
On Thursday, April 5, 2018, 10:09:18 AM PDT, Vlad Rozov
<vro...@apache.org> wrote:
Hi Paul,
Thank you, it is good to have a different angle view on Vectors,
Operators and Batches. For this thread, I'd suggest to keep focus on
DrillBuf design and implementation as the only gate for accessing raw
(direct) memory. It will be good for the community to agree on
1. Boundary checking (on/off based on a flag or assertions being
enabled/disabled, always on, always off, any other suggestions)
2. Ref count checking (delegate to netty or have a separate mechanism to
enable/disable, always on or off)
3. Usage of UDLE
4. Changing DrillBuf to follow Netty convention
5. Moving DrillBuf to a different package
Thank you,
Vlad
On 4/4/18 11:18, Paul Rogers wrote:
Hi Vlad,
Would be great to get insight from the original authors. Here ismy two cents as
a late comer who made extensive use of the classes in question.
Many of your questions are at the implementation level. It is worth looking at
the question from two other perspectives: history and design.
Historically, Drill adopted Netty for networking, and wisely looked for ways of
using the same buffers for both network transfer and internal operations to
avoid copies. Some overview is in [1]. In this view, a Drill vector is a
network buffer. Network buffers use the ByteBuffer protocol to serialize binary
values. DrillBuf follows that model for the most part. Because a ByteBuffer is
a low-level abstraction over a buffer, each operation must perform bounds
checks to ensure safe operation.
DrillBuf provides the ability to present a "view" of a slice of a larger underlying
buffer. For example, when reading data from a spill file, all data for all internal vectors is read
into a single buffer. For a nullable VarChar, for example, the buffer contains the bit vectors, the
offset vectors and the data vectors. The value vectors point to DrillBufs which point to a slice of
the underlying buffers. It is this layout (there are at least three different layouts) that makes
our "record batch sizer" so complex: the size of memory used is NOT the sum of the
DrillBufs.
Drill is a columnar system. So, the team introduced a typed "vector"
abstraction.Value vectors provide an abstraction that sweeps away the ByteBuffer heritage
and replaces it with a strongly typed, accessor/mutator structure that works in terms of
Drill data types and record counts. Vectors also understand the relationship between bit
vectors and data vectors, between offset vectors and data vectors, and so on.
Your question implies a desire to think about the future direction. Two things
to say. First, vectors themselves do not provide sufficient abstraction for the
needs of operators. As a result, operators become very complex, we must
generate large amounts of boiler-plate code, and we fix the same bugs over and
over. These issues are discussed at length in [2]. This is the motivation for
the result set reader and loader.
The row set abstractions encapsulate not just knowledge of a vector, but of the
entire batch. As a result, these abstractions know the number of records, know
the vector and batch size targets, and track vectors as they fill. One key
result is that these abstractions ensure that data is read or written within
the bounds of each buffer, eliminating the need for bounds checks on every
access.
The other consideration is memory management. Drill has a very complex, but surprisingly
robust, memory management system. However, it is based on a "malloc" model of
memory with operators negotiating among themselves (via the OUT_OF_MEMORY iterator
status) about who needs memory and who should release it. [2] discusses the limitations
of this system. As a result, we've been moving to a budget-based system in which each
fragment and operator is given a budget based on total available memory, and operators
use spilling to stay within the budget.
Memory fragmentation is a classic problem in malloc-based systems which strive
to operate at high memory utilization rates and which do not include memory
compaction. Drill is such a system. So, if this issue ever prevents Drill from
achieving maximum performance, we can consider the classic system used by
databases to solve this problem: fixed-size memory blocks.
If we were to move to fixed-size buffers, we'd want the row set and vector
abstractions to remain unchanged. We'd only want to replace DrillBuf with a new
block-based abstraction, perhaps with chaining (a vector may consist of a chain
of, say, 1 MB blocks.) The buffer slicing mechanism would become unnecessary,
as would the existing malloc-based allocator. Instead, data would be read,
written and held in buffers allocated from and returned to a buffer pool.
We may or may not ever make such a change. But, by considering this
possibility, we readily see that DrillBuf should be an implementation detail of
the higher-level abstractions and that operators should only use those
higher-level abstractions because doing so isolates operators from the details
of memory layout. This argument applies even more so to the abstractions below
DrillBuf: UDLE, Netty ByteBuf, ledgers and so on.
Said another way, even with the current system, we should be free to improve
DrillBuf on down with no impact to operator code because vectors and the row
set abstractions should be the only clients of DrillBuf.
In short, by understanding the history of the code, and agreeing upon the right
design abstractions, we can then make informed decisions about how best to
improve our low-level abstractions, including DrillBuf.
Thanks,
- Paul
[1] http://drill.apache.org/docs/value-vectors/
[2] https://github.com/paul-rogers/drill/wiki/Batch-Handling-Upgrades
On Wednesday, April 4, 2018, 10:34:18 AM PDT, Vlad Rozov
<vro...@apache.org> wrote:
I have several questions and concerns regarding DrillBuf usage, design
and implementation. There is a limited documentation available for the
subject (Java doc,
https://github.com/apache/drill/blob/master/exec/memory/base/src/main/java/org/apache/drill/exec/memory/README.md
and https://github.com/paul-rogers/drill/wiki/Memory-Management) and I
hope that a few members of the community may have more information.
What are the design goals behind DrillBuf? It seems like it is supposed
to be Drill access gate for direct byte buffers. How is it different
(for that goal) from UnsafeDirectLittleEndian? Both use
wrapper/delegation pattern, with DrillBuf delegating to
UnsafeDirectLittleEndian (not always) and UnsafeDirectLittleEndian
delegating to ByteBuf it wraps. Is it necessary to have both? Are there
any out of the box netty classes that already provide required
functionality? I guess that answer to the last question was "no" back
when DrillBuf and UnsafeDirectLittleEndian were introduced into Drill.
Is it still "no" for the latest netty release? What extra functionality
DrillBuf (and UnsafeDirectLittleEndian) provides on top of existing
netty classes?
As far as I can see from the source code, DrillBuf changes validation
(boundary and reference count checks) mechanism, making it optional
(compared to always enabled boundary checks inside netty) for get/set
Byte/Char/Short/Long/Float/Double. Is this a proper place to make
validation optional or the validation (or portion of the validation)
must be always on or off (there are different opinions, see
https://issues.apache.org/jira/browse/DRILL-6004,
https://issues.apache.org/jira/browse/DRILL-6202,
https://github.com/apache/drill/pull/1060 and
https://github.com/apache/drill/pull/1144)? Are there any performance
benchmark that justify or explain such behavior (if such benchmark does
not exist, are there any volunteer to do the benchmark)? My experience
is that the reference count check is significantly more expensive
compared to boundary checking and boundary checking adds tens of percent
to direct memory read when reading just a few bytes, so my vote is to
keep validation as optional with the ability to enable it for debug
purposes at run-time. What is the reason the same approach do not apply
to get/set Bytes and those methods are delegated to
UnsafeDirectLittleEndian that delegates it further?
Why DrillBuf reverses how AbstractByteBuf calls _get from get (and _set
from set), making _get to call get (_set to call set)? Why not to follow
a base class design patter?
Another question is usage of netty "io.netty.buffer" package for Drill
classes. Is this absolutely necessary? I don't think that netty
developers expect this and support semantic version compatibility for
package private classes/members.
Thank you,
Vlad
