blob: 6f0d61dd0279346d35181f12e8ac92ceda9d7b5a [file] [log] [blame]
<?xml version="1.0" encoding="UTF-8"?>
<!--
Licensed to the Apache Software Foundation (ASF) under one
or more contributor license agreements. See the NOTICE file
distributed with this work for additional information
regarding copyright ownership. The ASF licenses this file
to you under the Apache License, Version 2.0 (the
"License"); you may not use this file except in compliance
with the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing,
software distributed under the License is distributed on an
"AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
KIND, either express or implied. See the License for the
specific language governing permissions and limitations
under the License.
-->
<!DOCTYPE concept PUBLIC "-//OASIS//DTD DITA Concept//EN" "concept.dtd">
<concept id="scalability">
<title>Scalability Considerations for Impala</title>
<titlealts audience="PDF">
<navtitle>Scalability Considerations</navtitle>
</titlealts>
<prolog>
<metadata>
<data name="Category" value="Performance"/>
<data name="Category" value="Impala"/>
<data name="Category" value="Planning"/>
<data name="Category" value="Querying"/>
<data name="Category" value="Developers"/>
<data name="Category" value="Memory"/>
<data name="Category" value="Scalability"/>
<!-- Using domain knowledge about Impala, sizing, etc. to decide what to mark as 'Proof of Concept'. -->
<data name="Category" value="Proof of Concept"/>
</metadata>
</prolog>
<conbody>
<p>
This section explains how the size of your cluster and the volume of data influences SQL
performance and schema design for Impala tables. Typically, adding more cluster capacity
reduces problems due to memory limits or disk throughput. On the other hand, larger
clusters are more likely to have other kinds of scalability issues, such as a single slow
node that causes performance problems for queries.
</p>
<p outputclass="toc inpage"/>
<p conref="../shared/impala_common.xml#common/cookbook_blurb"/>
</conbody>
<concept audience="hidden" id="scalability_memory">
<title>Overview and Guidelines for Impala Memory Usage</title>
<prolog>
<metadata>
<data name="Category" value="Memory"/>
<data name="Category" value="Concepts"/>
<data name="Category" value="Best Practices"/>
<data name="Category" value="Guidelines"/>
</metadata>
</prolog>
<conbody>
<!--
Outline adapted from Alan Choi's "best practices" and/or "performance cookbook" papers.
-->
<codeblock>Memory Usage – the Basics
* Memory is used by:
* Hash join – RHS tables after decompression, filtering and projection
* Group by – proportional to the #groups
* Parquet writer buffer – 1GB per partition
* IO buffer (shared across queries)
* Metadata cache (no more than 1GB typically)
* Memory held and reused by later query
* Impala releases memory from time to time starting in 1.4.
Memory Usage – Estimating Memory Usage
* Use Explain Plan
* Requires statistics! Mem estimate without stats is meaningless.
* Reports per-host memory requirement for this cluster size.
* Re-run if you’ve re-sized the cluster!
[image of explain plan]
Memory Usage – Estimating Memory Usage
* EXPLAIN’s memory estimate issues
* Can be way off – much higher or much lower.
* group by’s estimate can be particularly off – when there’s a large number of group by columns.
* Mem estimate = NDV of group by column 1 * NDV of group by column 2 * ... NDV of group by column n
* Ignore EXPLAIN’s estimate if it’s too high! • Do your own estimate for group by
* GROUP BY mem usage = (total number of groups * size of each row) + (total number of groups * size of each row) / num node
Memory Usage – Finding Actual Memory Usage
* Search for “Per Node Peak Memory Usage” in the profile.
This is accurate. Use it for production capacity planning.
Memory Usage – Actual Memory Usage
* For complex queries, how do I know which part of my query is using too much memory?
* Use the ExecSummary from the query profile!
- But is that "Peak Mem" number aggregate or per-node?
[image of executive summary]
Memory Usage – Hitting Mem-limit
* Top causes (in order) of hitting mem-limit even when running a single query:
1. Lack of statistics
2. Lots of joins within a single query
3. Big-table joining big-table
4. Gigantic group by
Memory Usage – Hitting Mem-limit
Lack of stats
* Wrong join order, wrong join strategy, wrong insert strategy
* Explain Plan tells you that!
[image of explain plan]
* Fix: Compute Stats table
Memory Usage – Hitting Mem-limit
Lots of joins within a single query
* select...from fact, dim1, dim2,dim3,...dimN where ...
* Each dim tbl can fit in memory, but not all of them together
* As of Impala 1.4, Impala might choose the wrong plan – BROADCAST
FIX 1: use shuffle hint
select ... from fact join [shuffle] dim1 on ... join dim2 [shuffle] ...
FIX 2: pre-join the dim tables (if possible)
- How about an example to illustrate that technique?
* few join=&gt;better perf!
Memory Usage: Hitting Mem-limit
Big-table joining big-table
* Big-table (after decompression, filtering, and projection) is a table that is bigger than total cluster memory size.
* Impala 2.0 will do this (via disk-based join). Consider using Hive for now.
* (Advanced) For a simple query, you can try this advanced workaround – per-partition join
* Requires the partition key be part of the join key
select ... from BigTbl_A a join BigTbl_B b where a.part_key = b.part_key and a.part_key in (1,2,3)
union all
select ... from BigTbl_A a join BigTbl_B b where a.part_key = b.part_key and a.part_key in (4,5,6)
Memory Usage: Hitting Mem-limit
Gigantic group by
* The total number of distinct groups is huge, such as group by userid.
* Impala 2.0 will do this (via disk-based agg). Consider using Hive for now.
- Is this one of the cases where people were unhappy we recommended Hive?
* (Advanced) For a simple query, you can try this advanced workaround – per-partition agg
* Requires the partition key be part of the group by
select part_key, col1, col2, ...agg(..) from tbl where
part_key in (1,2,3)
Union all
Select part_key, col1, col2, ...agg(..) from tbl where
part_key in (4,5,6)
- But where's the GROUP BY in the preceding query? Need a real example.
Memory Usage: Additional Notes
* Use explain plan for estimate; use profile for accurate measure
* Data skew can use uneven memory usage
* Review previous common issues on out-of-memory
* Note: Even with disk-based joins, you'll want to review these steps to speed up queries and use memory more efficiently
</codeblock>
</conbody>
</concept>
<concept id="scalability_catalog">
<title>Impact of Many Tables or Partitions on Impala Catalog Performance and Memory Usage</title>
<conbody>
<p>
Because Hadoop I/O is optimized for reading and writing large files, Impala is optimized
for tables containing relatively few, large data files. Schemas containing thousands of
tables, or tables containing thousands of partitions, can encounter performance issues
during startup or during DDL operations such as <codeph>ALTER TABLE</codeph> statements.
</p>
<note type="important" rev="TSB-168">
<p>
Because of a change in the default heap size for the <cmdname>catalogd</cmdname>
daemon in <keyword
keyref="impala25_full"/> and higher, the following
procedure to increase the <cmdname>catalogd</cmdname> memory limit might be required
following an upgrade to <keyword keyref="impala25_full"/> even if not needed
previously.
</p>
</note>
<p conref="../shared/impala_common.xml#common/increase_catalogd_heap_size"
/>
</conbody>
</concept>
<concept rev="2.1.0" id="statestore_scalability">
<title>Scalability Considerations for the Impala Statestore</title>
<conbody>
<p>
Before <keyword keyref="impala21_full"/>, the statestore sent only one kind of message
to its subscribers. This message contained all updates for any topics that a subscriber
had subscribed to. It also served to let subscribers know that the statestore had not
failed, and conversely the statestore used the success of sending a heartbeat to a
subscriber to decide whether or not the subscriber had failed.
</p>
<p>
Combining topic updates and failure detection in a single message led to bottlenecks in
clusters with large numbers of tables, partitions, and HDFS data blocks. When the
statestore was overloaded with metadata updates to transmit, heartbeat messages were
sent less frequently, sometimes causing subscribers to time out their connection with
the statestore. Increasing the subscriber timeout and decreasing the frequency of
statestore heartbeats worked around the problem, but reduced responsiveness when the
statestore failed or restarted.
</p>
<p>
As of <keyword keyref="impala21_full"/>, the statestore now sends topic updates and
heartbeats in separate messages. This allows the statestore to send and receive a steady
stream of lightweight heartbeats, and removes the requirement to send topic updates
according to a fixed schedule, reducing statestore network overhead.
</p>
<p>
The statestore now has the following relevant configuration flags for the
<cmdname>statestored</cmdname> daemon:
</p>
<dl>
<dlentry id="statestore_num_update_threads">
<dt>
<codeph>-statestore_num_update_threads</codeph>
</dt>
<dd>
The number of threads inside the statestore dedicated to sending topic updates. You
should not typically need to change this value.
<p>
<b>Default:</b> 10
</p>
</dd>
</dlentry>
<dlentry id="statestore_update_frequency_ms">
<dt>
<codeph>-statestore_update_frequency_ms</codeph>
</dt>
<dd>
The frequency, in milliseconds, with which the statestore tries to send topic
updates to each subscriber. This is a best-effort value; if the statestore is unable
to meet this frequency, it sends topic updates as fast as it can. You should not
typically need to change this value.
<p>
<b>Default:</b> 2000
</p>
</dd>
</dlentry>
<dlentry id="statestore_num_heartbeat_threads">
<dt>
<codeph>-statestore_num_heartbeat_threads</codeph>
</dt>
<dd>
The number of threads inside the statestore dedicated to sending heartbeats. You
should not typically need to change this value.
<p>
<b>Default:</b> 10
</p>
</dd>
</dlentry>
<dlentry id="statestore_heartbeat_frequency_ms">
<dt>
<codeph>-statestore_heartbeat_frequency_ms</codeph>
</dt>
<dd>
The frequency, in milliseconds, with which the statestore tries to send heartbeats
to each subscriber. This value should be good for large catalogs and clusters up to
approximately 150 nodes. Beyond that, you might need to increase this value to make
the interval longer between heartbeat messages.
<p>
<b>Default:</b> 1000 (one heartbeat message every second)
</p>
</dd>
</dlentry>
<dlentry id="statestore_heartbeat_tcp_timeout_seconds">
<dt>
<codeph>-statestore_heartbeat_tcp_timeout_seconds</codeph>
</dt>
<dd>
The time after which a heartbeat RPC to a subscriber will timeout. This setting
protects against badly hung machines that are not able to respond to the heartbeat
RPC in short order. Increase this if there are intermittent heartbeat RPC timeouts
shown in statestore's log. You can reference the max value of
"statestore.priority-topic-update-durations" metric on statestore to get a
reasonable value. Note that priority topic updates are assumed to be small amounts
of data that take a small amount of time to process (similar to the heartbeat
complexity).
<p>
<b>Default:</b> 3
</p>
</dd>
</dlentry>
<dlentry id="statestore_max_missed_heartbeats">
<dt>
<codeph>-statestore_max_missed_heartbeats</codeph>
</dt>
<dd>
Maximum number of consecutive heartbeat messages an impalad can miss before being
declared failed by the statestore. You should not typically need to change this
value.
<p>
<b>Default:</b> 10
</p>
</dd>
</dlentry>
<dlentry id="statestore_subscriber_timeout_secs">
<dt>
<codeph>-statestore_subscriber_timeout_secs</codeph>
</dt>
<dd>
The amount of time (in seconds) that may elapse before the connection with the
statestore is considered lost by subscribers (impalad/catalogd). Impalad will
reregister itself to statestore, which may cause its absence in the next round of
cluster membership update. This will cause query failures like "Cancelled due to
unreachable impalad(s)". The value of this flag should be comparable to
<codeph>
(statestore_heartbeat_frequency_ms / 1000 + statestore_heartbeat_tcp_timeout_seconds)
* statestore_max_missed_heartbeats</codeph>,
so subscribers won't reregister themselves too early and allow statestore to
resend heartbeats. You can also reference the max value of
"statestore-subscriber.heartbeat-interval-time" metrics on impalads to get a
reasonable value.
<p>
<b>Default:</b> 30
</p>
</dd>
</dlentry>
</dl>
<p>
If it takes a very long time for a cluster to start up, and
<cmdname>impala-shell</cmdname> consistently displays <codeph>This Impala daemon is not
ready to accept user requests</codeph>, the statestore might be taking too long to send
the entire catalog topic to the cluster. In this case, consider adding
<codeph>--load_catalog_in_background=false</codeph> to your catalog service
configuration. This setting stops the statestore from loading the entire catalog into
memory at cluster startup. Instead, metadata for each table is loaded when the table is
accessed for the first time.
</p>
</conbody>
</concept>
<concept id="scalability_buffer_pool" rev="2.10.0 IMPALA-3200">
<title>Effect of Buffer Pool on Memory Usage (<keyword keyref="impala210"/> and higher)</title>
<conbody>
<p>
The buffer pool feature, available in <keyword keyref="impala210"/> and higher, changes
the way Impala allocates memory during a query. Most of the memory needed is reserved at
the beginning of the query, avoiding cases where a query might run for a long time
before failing with an out-of-memory error. The actual memory estimates and memory
buffers are typically smaller than before, so that more queries can run concurrently or
process larger volumes of data than previously.
</p>
<p>
The buffer pool feature includes some query options that you can fine-tune:
<xref keyref="buffer_pool_limit"/>,
<xref
keyref="default_spillable_buffer_size"/>,
<xref keyref="max_row_size"
/>, and <xref keyref="min_spillable_buffer_size"/>.
</p>
<p>
Most of the effects of the buffer pool are transparent to you as an Impala user. Memory
use during spilling is now steadier and more predictable, instead of increasing rapidly
as more data is spilled to disk. The main change from a user perspective is the need to
increase the <codeph>MAX_ROW_SIZE</codeph> query option setting when querying tables
with columns containing long strings, many columns, or other combinations of factors
that produce very large rows. If Impala encounters rows that are too large to process
with the default query option settings, the query fails with an error message suggesting
to increase the <codeph>MAX_ROW_SIZE</codeph> setting.
</p>
</conbody>
</concept>
<concept audience="hidden" id="scalability_cluster_size">
<title>Scalability Considerations for Impala Cluster Size and Topology</title>
<conbody>
<p/>
</conbody>
</concept>
<concept audience="hidden" id="concurrent_connections">
<title>Scaling the Number of Concurrent Connections</title>
<conbody>
<p/>
</conbody>
</concept>
<concept rev="2.0.0" id="spill_to_disk">
<title>SQL Operations that Spill to Disk</title>
<conbody>
<p>
Certain memory-intensive operations write temporary data to disk (known as
<term>spilling</term> to disk) when Impala is close to exceeding its memory limit on a
particular host.
</p>
<p>
The result is a query that completes successfully, rather than failing with an
out-of-memory error. The tradeoff is decreased performance due to the extra disk I/O to
write the temporary data and read it back in. The slowdown could be potentially be
significant. Thus, while this feature improves reliability, you should optimize your
queries, system parameters, and hardware configuration to make this spilling a rare
occurrence.
</p>
<note rev="2.10.0 IMPALA-3200">
<p>
In <keyword keyref="impala210"/> and higher, also see
<xref
keyref="scalability_buffer_pool"/> for changes to Impala memory
allocation that might change the details of which queries spill to disk, and how much
memory and disk space is involved in the spilling operation.
</p>
</note>
<p>
<b>What kinds of queries might spill to disk:</b>
</p>
<p>
Several SQL clauses and constructs require memory allocations that could activat the
spilling mechanism:
</p>
<ul>
<li>
<p>
when a query uses a <codeph>GROUP BY</codeph> clause for columns with millions or
billions of distinct values, Impala keeps a similar number of temporary results in
memory, to accumulate the aggregate results for each value in the group.
</p>
</li>
<li>
<p>
When large tables are joined together, Impala keeps the values of the join columns
from one table in memory, to compare them to incoming values from the other table.
</p>
</li>
<li>
<p>
When a large result set is sorted by the <codeph>ORDER BY</codeph> clause, each node
sorts its portion of the result set in memory.
</p>
</li>
<li>
<p>
The <codeph>DISTINCT</codeph> and <codeph>UNION</codeph> operators build in-memory
data structures to represent all values found so far, to eliminate duplicates as the
query progresses.
</p>
</li>
<!-- JIRA still in open state as of 5.8 / 2.6, commenting out.
<li>
<p rev="IMPALA-3471">
In <keyword keyref="impala26_full"/> and higher, <term>top-N</term> queries (those with
<codeph>ORDER BY</codeph> and <codeph>LIMIT</codeph> clauses) can also spill.
Impala allocates enough memory to hold as many rows as specified by the <codeph>LIMIT</codeph>
clause, plus enough memory to hold as many rows as specified by any <codeph>OFFSET</codeph> clause.
</p>
</li>
-->
</ul>
<p
conref="../shared/impala_common.xml#common/spill_to_disk_vs_dynamic_partition_pruning"/>
<p>
<b>How Impala handles scratch disk space for spilling:</b>
</p>
<p rev="obwl"
conref="../shared/impala_common.xml#common/order_by_scratch_dir"/>
<p>
<b>Memory usage for SQL operators:</b>
</p>
<p rev="2.10.0 IMPALA-3200">
In <keyword keyref="impala210_full"/> and higher, the way SQL operators such as
<codeph>GROUP BY</codeph>, <codeph>DISTINCT</codeph>, and joins, transition between
using additional memory or activating the spill-to-disk feature is changed. The memory
required to spill to disk is reserved up front, and you can examine it in the
<codeph>EXPLAIN</codeph> plan when the <codeph>EXPLAIN_LEVEL</codeph> query option is
set to 2 or higher.
</p>
<p>
The infrastructure of the spilling feature affects the way the affected SQL operators,
such as <codeph>GROUP BY</codeph>, <codeph>DISTINCT</codeph>, and joins, use memory. On
each host that participates in the query, each such operator in a query requires memory
to store rows of data and other data structures. Impala reserves a certain amount of
memory up front for each operator that supports spill-to-disk that is sufficient to
execute the operator. If an operator accumulates more data than can fit in the reserved
memory, it can either reserve more memory to continue processing data in memory or start
spilling data to temporary scratch files on disk. Thus, operators with spill-to-disk
support can adapt to different memory constraints by using however much memory is
available to speed up execution, yet tolerate low memory conditions by spilling data to
disk.
</p>
<p>
The amount data depends on the portion of the data being handled by that host, and thus
the operator may end up consuming different amounts of memory on different hosts.
</p>
<!--
<p>
The infrastructure of the spilling feature affects the way the affected SQL operators, such as
<codeph>GROUP BY</codeph>, <codeph>DISTINCT</codeph>, and joins, use memory.
On each host that participates in the query, each such operator in a query accumulates memory
while building the data structure to process the aggregation or join operation. The amount
of memory used depends on the portion of the data being handled by that host, and thus might
be different from one host to another. When the amount of memory being used for the operator
on a particular host reaches a threshold amount, Impala reserves an additional memory buffer
to use as a work area in case that operator causes the query to exceed the memory limit for
that host. After allocating the memory buffer, the memory used by that operator remains
essentially stable or grows only slowly, until the point where the memory limit is reached
and the query begins writing temporary data to disk.
</p>
<p rev="2.2.0">
Prior to Impala 2.2, the extra memory buffer for an operator that might spill to disk
was allocated when the data structure used by the applicable SQL operator reaches 16 MB in size,
and the memory buffer itself was 512 MB. In Impala 2.2, these values are halved: the threshold value
is 8 MB and the memory buffer is 256 MB. <ph rev="2.3.0">In <keyword keyref="impala23_full"/> and higher, the memory for the buffer
is allocated in pieces, only as needed, to avoid sudden large jumps in memory usage.</ph> A query that uses
multiple such operators might allocate multiple such memory buffers, as the size of the data structure
for each operator crosses the threshold on a particular host.
</p>
<p>
Therefore, a query that processes a relatively small amount of data on each host would likely
never reach the threshold for any operator, and would never allocate any extra memory buffers. A query
that did process millions of groups, distinct values, join keys, and so on might cross the threshold,
causing its memory requirement to rise suddenly and then flatten out. The larger the cluster, less data is processed
on any particular host, thus reducing the chance of requiring the extra memory allocation.
</p>
-->
<p>
<b>Added in:</b> This feature was added to the <codeph>ORDER BY</codeph> clause in
Impala 1.4. This feature was extended to cover join queries, aggregation functions, and
analytic functions in Impala 2.0. The size of the memory work area required by each
operator that spills was reduced from 512 megabytes to 256 megabytes in Impala 2.2.
<ph
rev="2.10.0 IMPALA-3200">The spilling mechanism was reworked to take
advantage of the Impala buffer pool feature and be more predictable and stable in
<keyword keyref="impala210_full"/>.</ph>
</p>
<p>
<b>Avoiding queries that spill to disk:</b>
</p>
<p>
Because the extra I/O can impose significant performance overhead on these types of
queries, try to avoid this situation by using the following steps:
</p>
<ol>
<li>
Detect how often queries spill to disk, and how much temporary data is written. Refer
to the following sources:
<ul>
<li>
The output of the <codeph>PROFILE</codeph> command in the
<cmdname>impala-shell</cmdname> interpreter. This data shows the memory usage for
each host and in total across the cluster. The <codeph>WriteIoBytes</codeph>
counter reports how much data was written to disk for each operator during the
query. (In <keyword
keyref="impala29_full"/>, the counter was
named <codeph>ScratchBytesWritten</codeph>; in
<keyword
keyref="impala28_full"/> and earlier, it was named
<codeph>BytesWritten</codeph>.)
</li>
<li>
The <uicontrol>Queries</uicontrol> tab in the Impala debug web user interface.
Select the query to examine and click the corresponding
<uicontrol>Profile</uicontrol> link. This data breaks down the memory usage for a
single host within the cluster, the host whose web interface you are connected to.
</li>
</ul>
</li>
<li>
Use one or more techniques to reduce the possibility of the queries spilling to disk:
<ul>
<li>
Increase the Impala memory limit if practical, for example, if you can increase
the available memory by more than the amount of temporary data written to disk on
a particular node. Remember that in Impala 2.0 and later, you can issue
<codeph>SET MEM_LIMIT</codeph> as a SQL statement, which lets you fine-tune the
memory usage for queries from JDBC and ODBC applications.
</li>
<li>
Increase the number of nodes in the cluster, to increase the aggregate memory
available to Impala and reduce the amount of memory required on each node.
</li>
<li>
Add more memory to the hosts running Impala daemons.
</li>
<li>
On a cluster with resources shared between Impala and other Hadoop components, use
resource management features to allocate more memory for Impala. See
<xref
href="impala_resource_management.xml#resource_management"/>
for details.
</li>
<li>
If the memory pressure is due to running many concurrent queries rather than a few
memory-intensive ones, consider using the Impala admission control feature to
lower the limit on the number of concurrent queries. By spacing out the most
resource-intensive queries, you can avoid spikes in memory usage and improve
overall response times. See
<xref
href="impala_admission.xml#admission_control"/> for details.
</li>
<li>
Tune the queries with the highest memory requirements, using one or more of the
following techniques:
<ul>
<li>
Run the <codeph>COMPUTE STATS</codeph> statement for all tables involved in
large-scale joins and aggregation queries.
</li>
<li>
Minimize your use of <codeph>STRING</codeph> columns in join columns. Prefer
numeric values instead.
</li>
<li>
Examine the <codeph>EXPLAIN</codeph> plan to understand the execution strategy
being used for the most resource-intensive queries. See
<xref href="impala_explain_plan.xml#perf_explain"
/> for
details.
</li>
<li>
If Impala still chooses a suboptimal execution strategy even with statistics
available, or if it is impractical to keep the statistics up to date for huge
or rapidly changing tables, add hints to the most resource-intensive queries
to select the right execution strategy. See
<xref
href="impala_hints.xml#hints"/> for details.
</li>
</ul>
</li>
<li>
If your queries experience substantial performance overhead due to spilling,
enable the <codeph>DISABLE_UNSAFE_SPILLS</codeph> query option. This option
prevents queries whose memory usage is likely to be exorbitant from spilling to
disk. See
<xref
href="impala_disable_unsafe_spills.xml#disable_unsafe_spills"/>
for details. As you tune problematic queries using the preceding steps, fewer and
fewer will be cancelled by this option setting.
</li>
</ul>
</li>
</ol>
<p>
<b>Testing performance implications of spilling to disk:</b>
</p>
<p>
To artificially provoke spilling, to test this feature and understand the performance
implications, use a test environment with a memory limit of at least 2 GB. Issue the
<codeph>SET</codeph> command with no arguments to check the current setting for the
<codeph>MEM_LIMIT</codeph> query option. Set the query option
<codeph>DISABLE_UNSAFE_SPILLS=true</codeph>. This option limits the spill-to-disk
feature to prevent runaway disk usage from queries that are known in advance to be
suboptimal. Within <cmdname>impala-shell</cmdname>, run a query that you expect to be
memory-intensive, based on the criteria explained earlier. A self-join of a large table
is a good candidate:
</p>
<codeblock>select count(*) from big_table a join big_table b using (column_with_many_values);
</codeblock>
<p>
Issue the <codeph>PROFILE</codeph> command to get a detailed breakdown of the memory
usage on each node during the query.
<!--
The crucial part of the profile output concerning memory is the <codeph>BlockMgr</codeph>
portion. For example, this profile shows that the query did not quite exceed the memory limit.
-->
</p>
<!-- Commenting out because now stale due to changes from the buffer pool (IMPALA-3200).
To do: Revisit these details later if indicated by user feedback.
<codeblock>BlockMgr:
- BlockWritesIssued: 1
- BlockWritesOutstanding: 0
- BlocksCreated: 24
- BlocksRecycled: 1
- BufferedPins: 0
- MaxBlockSize: 8.00 MB (8388608)
<b>- MemoryLimit: 200.00 MB (209715200)</b>
<b>- PeakMemoryUsage: 192.22 MB (201555968)</b>
- TotalBufferWaitTime: 0ns
- TotalEncryptionTime: 0ns
- TotalIntegrityCheckTime: 0ns
- TotalReadBlockTime: 0ns
</codeblock>
<p>
In this case, because the memory limit was already below any recommended value, I increased the volume of
data for the query rather than reducing the memory limit any further.
</p>
-->
<p>
Set the <codeph>MEM_LIMIT</codeph> query option to a value that is smaller than the peak
memory usage reported in the profile output. Now try the memory-intensive query again.
</p>
<p>
Check if the query fails with a message like the following:
</p>
<codeblock>WARNINGS: Spilling has been disabled for plans that do not have stats and are not hinted
to prevent potentially bad plans from using too many cluster resources. Compute stats on
these tables, hint the plan or disable this behavior via query options to enable spilling.
</codeblock>
<p>
If so, the query could have consumed substantial temporary disk space, slowing down so
much that it would not complete in any reasonable time. Rather than rely on the
spill-to-disk feature in this case, issue the <codeph>COMPUTE STATS</codeph> statement
for the table or tables in your sample query. Then run the query again, check the peak
memory usage again in the <codeph>PROFILE</codeph> output, and adjust the memory limit
again if necessary to be lower than the peak memory usage.
</p>
<p>
At this point, you have a query that is memory-intensive, but Impala can optimize it
efficiently so that the memory usage is not exorbitant. You have set an artificial
constraint through the <codeph>MEM_LIMIT</codeph> option so that the query would
normally fail with an out-of-memory error. But the automatic spill-to-disk feature means
that the query should actually succeed, at the expense of some extra disk I/O to read
and write temporary work data.
</p>
<p>
Try the query again, and confirm that it succeeds. Examine the <codeph>PROFILE</codeph>
output again. This time, look for lines of this form:
</p>
<codeblock>- SpilledPartitions: <varname>N</varname>
</codeblock>
<p>
If you see any such lines with <varname>N</varname> greater than 0, that indicates the
query would have failed in Impala releases prior to 2.0, but now it succeeded because of
the spill-to-disk feature. Examine the total time taken by the
<codeph>AGGREGATION_NODE</codeph> or other query fragments containing non-zero
<codeph>SpilledPartitions</codeph> values. Compare the times to similar fragments that
did not spill, for example in the <codeph>PROFILE</codeph> output when the same query is
run with a higher memory limit. This gives you an idea of the performance penalty of the
spill operation for a particular query with a particular memory limit. If you make the
memory limit just a little lower than the peak memory usage, the query only needs to
write a small amount of temporary data to disk. The lower you set the memory limit, the
more temporary data is written and the slower the query becomes.
</p>
<p>
Now repeat this procedure for actual queries used in your environment. Use the
<codeph>DISABLE_UNSAFE_SPILLS</codeph> setting to identify cases where queries used more
memory than necessary due to lack of statistics on the relevant tables and columns, and
issue <codeph>COMPUTE STATS</codeph> where necessary.
</p>
<p>
<b>When to use DISABLE_UNSAFE_SPILLS:</b>
</p>
<p>
You might wonder, why not leave <codeph>DISABLE_UNSAFE_SPILLS</codeph> turned on all the
time. Whether and how frequently to use this option depends on your system environment
and workload.
</p>
<p>
<codeph>DISABLE_UNSAFE_SPILLS</codeph> is suitable for an environment with ad hoc
queries whose performance characteristics and memory usage are not known in advance. It
prevents <q>worst-case scenario</q> queries that use large amounts of memory
unnecessarily. Thus, you might turn this option on within a session while developing new
SQL code, even though it is turned off for existing applications.
</p>
<p>
Organizations where table and column statistics are generally up-to-date might leave
this option turned on all the time, again to avoid worst-case scenarios for untested
queries or if a problem in the ETL pipeline results in a table with no statistics.
Turning on <codeph>DISABLE_UNSAFE_SPILLS</codeph> lets you <q>fail fast</q> in this case
and immediately gather statistics or tune the problematic queries.
</p>
<p>
Some organizations might leave this option turned off. For example, you might have
tables large enough that the <codeph>COMPUTE STATS</codeph> takes substantial time to
run, making it impractical to re-run after loading new data. If you have examined the
<codeph>EXPLAIN</codeph> plans of your queries and know that they are operating
efficiently, you might leave <codeph>DISABLE_UNSAFE_SPILLS</codeph> turned off. In that
case, you know that any queries that spill will not go overboard with their memory
consumption.
</p>
</conbody>
</concept>
<concept id="complex_query">
<title>Limits on Query Size and Complexity</title>
<conbody>
<p>
There are hardcoded limits on the maximum size and complexity of queries. Currently, the
maximum number of expressions in a query is 2000. You might exceed the limits with large
or deeply nested queries produced by business intelligence tools or other query
generators.
</p>
<p>
If you have the ability to customize such queries or the query generation logic that
produces them, replace sequences of repetitive expressions with single operators such as
<codeph>IN</codeph> or <codeph>BETWEEN</codeph> that can represent multiple values or
ranges. For example, instead of a large number of <codeph>OR</codeph> clauses:
</p>
<codeblock>WHERE val = 1 OR val = 2 OR val = 6 OR val = 100 ...
</codeblock>
<p>
use a single <codeph>IN</codeph> clause:
</p>
<codeblock>WHERE val IN (1,2,6,100,...)</codeblock>
</conbody>
</concept>
<concept id="scalability_io">
<title>Scalability Considerations for Impala I/O</title>
<conbody>
<p>
Impala parallelizes its I/O operations aggressively, therefore the more disks you can
attach to each host, the better. Impala retrieves data from disk so quickly using bulk
read operations on large blocks, that most queries are CPU-bound rather than I/O-bound.
</p>
<p>
Because the kind of sequential scanning typically done by Impala queries does not
benefit much from the random-access capabilities of SSDs, spinning disks typically
provide the most cost-effective kind of storage for Impala data, with little or no
performance penalty as compared to SSDs.
</p>
<p>
Resource management features such as YARN, Llama, and admission control typically
constrain the amount of memory, CPU, or overall number of queries in a high-concurrency
environment. Currently, there is no throttling mechanism for Impala I/O.
</p>
</conbody>
</concept>
<concept id="big_tables">
<title>Scalability Considerations for Table Layout</title>
<conbody>
<p>
Due to the overhead of retrieving and updating table metadata in the metastore database,
try to limit the number of columns in a table to a maximum of approximately 2000.
Although Impala can handle wider tables than this, the metastore overhead can become
significant, leading to query performance that is slower than expected based on the
actual data volume.
</p>
<p>
To minimize overhead related to the metastore database and Impala query planning, try to
limit the number of partitions for any partitioned table to a few tens of thousands.
</p>
<p rev="IMPALA-5309">
If the volume of data within a table makes it impractical to run exploratory queries,
consider using the <codeph>TABLESAMPLE</codeph> clause to limit query processing to only
a percentage of data within the table. This technique reduces the overhead for query
startup, I/O to read the data, and the amount of network, CPU, and memory needed to
process intermediate results during the query. See <xref keyref="tablesample"/> for
details.
</p>
</conbody>
</concept>
<concept rev="" id="kerberos_overhead_cluster_size">
<title>Kerberos-Related Network Overhead for Large Clusters</title>
<conbody>
<p>
When Impala starts up, or after each <codeph>kinit</codeph> refresh, Impala sends a
number of simultaneous requests to the KDC. For a cluster with 100 hosts, the KDC might
be able to process all the requests within roughly 5 seconds. For a cluster with 1000
hosts, the time to process the requests would be roughly 500 seconds. Impala also makes
a number of DNS requests at the same time as these Kerberos-related requests.
</p>
<p>
While these authentication requests are being processed, any submitted Impala queries
will fail. During this period, the KDC and DNS may be slow to respond to requests from
components other than Impala, so other secure services might be affected temporarily.
</p>
<p>
In <keyword keyref="impala212_full"/> or earlier, to reduce the frequency of the
<codeph>kinit</codeph> renewal that initiates a new set of authentication requests,
increase the <codeph>kerberos_reinit_interval</codeph> configuration setting for the
<codeph>impalad</codeph> daemons. Currently, the default is 60 minutes. Consider using a
higher value such as 360 (6 hours).
</p>
<p>
The <codeph>kerberos_reinit_interval</codeph> configuration setting is removed in
<keyword keyref="impala30_full"/>, and the above step is no longer needed.
</p>
</conbody>
</concept>
<concept id="scalability_hotspots" rev="2.5.0 IMPALA-2696">
<title>Avoiding CPU Hotspots for HDFS Cached Data</title>
<conbody>
<p>
You can use the HDFS caching feature, described in
<xref
href="impala_perf_hdfs_caching.xml#hdfs_caching"/>, with Impala to
reduce I/O and memory-to-memory copying for frequently accessed tables or partitions.
</p>
<p>
In the early days of this feature, you might have found that enabling HDFS caching
resulted in little or no performance improvement, because it could result in
<q>hotspots</q>: instead of the I/O to read the table data being parallelized across the
cluster, the I/O was reduced but the CPU load to process the data blocks might be
concentrated on a single host.
</p>
<p>
To avoid hotspots, include the <codeph>WITH REPLICATION</codeph> clause with the
<codeph>CREATE TABLE</codeph> or <codeph>ALTER TABLE</codeph> statements for tables that
use HDFS caching. This clause allows more than one host to cache the relevant data
blocks, so the CPU load can be shared, reducing the load on any one host. See
<xref
href="impala_create_table.xml#create_table"/> and
<xref
href="impala_alter_table.xml#alter_table"/> for details.
</p>
<p>
Hotspots with high CPU load for HDFS cached data could still arise in some cases, due to
the way that Impala schedules the work of processing data blocks on different hosts. In
<keyword keyref="impala25_full"/> and higher, scheduling improvements mean that the work
for HDFS cached data is divided better among all the hosts that have cached replicas for
a particular data block. When more than one host has a cached replica for a data block,
Impala assigns the work of processing that block to whichever host has done the least
work (in terms of number of bytes read) for the current query. If hotspots persist even
with this load-based scheduling algorithm, you can enable the query option
<codeph>SCHEDULE_RANDOM_REPLICA=TRUE</codeph> to further distribute the CPU load. This
setting causes Impala to randomly pick a host to process a cached data block if the
scheduling algorithm encounters a tie when deciding which host has done the least work.
</p>
</conbody>
</concept>
<concept id="scalability_file_handle_cache" rev="2.10.0 IMPALA-4623">
<title>Scalability Considerations for File Handle Caching</title>
<conbody>
<p>
One scalability aspect that affects heavily loaded clusters is the load on the metadata
layer from looking up the details as each file is opened. On HDFS, that can lead to
increased load on the NameNode, and on S3, this can lead to an excessive number of S3
metadata requests. For example, a query that does a full table scan on a partitioned
table may need to read thousands of partitions, each partition containing multiple data
files. Accessing each column of a Parquet file also involves a separate <q>open</q>
call, further increasing the load on the NameNode. High NameNode overhead can add
startup time (that is, increase latency) to Impala queries, and reduce overall
throughput for non-Impala workloads that also require accessing HDFS files.
</p>
<p>
You can reduce the number of calls made to your file system's metadata layer by enabling
the file handle caching feature. Data files that are accessed by different queries, or
even multiple times within the same query, can be accessed without a new <q>open</q>
call and without fetching the file details multiple times.
</p>
<p>
Impala supports file handle caching for the following file systems:
<ul>
<li>
HDFS in <keyword keyref="impala210_full"/> and higher
<p>
In Impala 3.2 and higher, file handle caching also applies to remote HDFS file
handles. This is controlled by the <codeph>cache_remote_file_handles</codeph> flag
for an <codeph>impalad</codeph>. It is recommended that you use the default value
of <codeph>true</codeph> as this caching prevents your NameNode from overloading
when your cluster has many remote HDFS reads.
</p>
</li>
<li>
S3 in <keyword keyref="impala33_full"/> and higher
<p>
The <codeph>cache_s3_file_handles</codeph> <codeph>impalad</codeph> flag controls
the S3 file handle caching. The feature is enabled by default with the flag set to
<codeph>true</codeph>.
</p>
</li>
</ul>
</p>
<p>
The feature is enabled by default with 20,000 file handles to be cached. To change the
value, set the configuration option <codeph>max_cached_file_handles</codeph> to a
non-zero value for each <cmdname>impalad</cmdname> daemon. From the initial default
value of 20000, adjust upward if NameNode request load is still significant, or downward
if it is more important to reduce the extra memory usage on each host. Each cache entry
consumes 6 KB, meaning that caching 20,000 file handles requires up to 120 MB on each
Impala executor. The exact memory usage varies depending on how many file handles have
actually been cached; memory is freed as file handles are evicted from the cache.
</p>
<p>
If a manual operation moves a file to the trashcan while the file handle is cached,
Impala still accesses the contents of that file. This is a change from prior behavior.
Previously, accessing a file that was in the trashcan would cause an error. This
behavior only applies to non-Impala methods of removing files, not the Impala mechanisms
such as <codeph>TRUNCATE TABLE</codeph> or <codeph>DROP TABLE</codeph>.
</p>
<p>
If files are removed, replaced, or appended by operations outside of Impala, the way to
bring the file information up to date is to run the <codeph>REFRESH</codeph> statement
on the table.
</p>
<p>
File handle cache entries are evicted as the cache fills up, or based on a timeout
period when they have not been accessed for some time.
</p>
<p>
To evaluate the effectiveness of file handle caching for a particular workload, issue
the <codeph>PROFILE</codeph> statement in <cmdname>impala-shell</cmdname> or examine
query profiles in the Impala Web UI. Look for the ratio of
<codeph>CachedFileHandlesHitCount</codeph> (ideally, should be high) to
<codeph>CachedFileHandlesMissCount</codeph> (ideally, should be low). Before starting
any evaluation, run several representative queries to <q>warm up</q> the cache because
the first time each data file is accessed is always recorded as a cache miss.
</p>
<p>
To see metrics about file handle caching for each <cmdname>impalad</cmdname> instance,
examine the following fields on the <uicontrol>/metrics</uicontrol> page in the Impala
Web UI:
</p>
<ul>
<li>
<uicontrol>impala-server.io.mgr.cached-file-handles-miss-count</uicontrol>
</li>
<li>
<uicontrol>impala-server.io.mgr.num-cached-file-handles</uicontrol>
</li>
</ul>
</conbody>
</concept>
</concept>