[
https://issues.apache.org/jira/browse/PHOENIX-7106?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Viraj Jasani updated PHOENIX-7106:
----------------------------------
Description:
HBase scanner interface expects server to perform scan of the cells from HFile
or Block cache and return consistent data i.e. rowkey of the cells returned
should stay in the range of the scan boundaries. When a region moves and
scanner needs reset, or if the current row is too large and the server returns
partial row, the subsequent scanner#next is supposed to return remaining cells.
When this happens, cell rowkeys returned by servers i.e. any coprocessors is
expected to be in the scan boundary range so that server can reliably perform
its validation and return remaining cells as expected.
Phoenix client initiates serial or parallel scans from the aggregators based on
the region boundaries and the scan boundaries are sometimes adjusted based on
where optimizer provided key ranges, to include tenant boundaries, salt
boundaries etc. After the client opens the scanner and performs scan operation,
some of the coprocs return invalid rowkey for the following cases:
# Grouped aggregate queries
# Some Ungrouped aggregate queries
# Offset queries
# Dummy cells returned with empty rowkey
# Update statistics queries
# Uncovered Index queries
# Ordered results at server side
# ORDER BY DESC on rowkey
# Global Index read-repair
# Paging region scanner with HBase scanner reopen
# ORDER BY on non-pk column(s) with/without paging
# GROUP BY on non-pk column(s) with/without paging
Since many of these cases return reserved rowkeys, they are likely not going to
match scan or region boundaries. It has potential to cause data integrity
issues in certain scenarios as explained above. Empty rowkey returned by server
can be treated as end of the region scan by HBase client.
With the paging feature enabled, if the page size is kept low, we have higher
chances of scanners returning dummy cell, resulting in increased num of RPC
calls for better latency and timeouts. We should return only valid rowkey in
the scan range for all the cases where we perform above mentioned operations
like complex aggregate or offset queries etc.
was:
HBase scanner interface expects server to perform scan of the cells from HFile
or Block cache and return consistent data i.e. rowkey of the cells returned
should stay in the range of the scan boundaries. When a region moves and
scanner needs reset, or if the current row is too large and the server returns
partial row, the subsequent scanner#next is supposed to return remaining cells.
When this happens, cell rowkeys returned by servers i.e. any coprocessors is
expected to be in the scan boundary range so that server can reliably perform
its validation and return remaining cells as expected.
Phoenix client initiates serial or parallel scans from the aggregators based on
the region boundaries and the scan boundaries are sometimes adjusted based on
where optimizer provided key ranges, to include tenant boundaries, salt
boundaries etc. After the client opens the scanner and performs scan operation,
some of the coprocs return invalid rowkey for the following cases:
# Grouped aggregate queries
# Ungrouped aggregate queries (not all of them)
# Offset queries
# Some dummy cells returned with empty rowkey
# Update statistics queries
# Local indexes
Since many of these cases return reserved rowkeys, they are likely not going to
match scan or region boundaries. It has potential to cause data integrity
issues in certain scenarios as explained above. Empty rowkey returned by server
can be treated as end of the region scan by HBase client.
With the paging feature enabled, if the page size is kept low, we have higher
chances of scanners returning dummy cell, resulting in increased num of RPC
calls for better latency and timeouts. We should return only valid rowkey in
the scan range for all the cases where we perform above mentioned operations
like complex aggregate or offset queries etc.
> Data Integrity issues due to invalid rowkeys returned by various coprocessors
> -----------------------------------------------------------------------------
>
> Key: PHOENIX-7106
> URL: https://issues.apache.org/jira/browse/PHOENIX-7106
> Project: Phoenix
> Issue Type: Improvement
> Reporter: Viraj Jasani
> Assignee: Viraj Jasani
> Priority: Critical
> Fix For: 5.2.0, 5.1.4
>
>
> HBase scanner interface expects server to perform scan of the cells from
> HFile or Block cache and return consistent data i.e. rowkey of the cells
> returned should stay in the range of the scan boundaries. When a region moves
> and scanner needs reset, or if the current row is too large and the server
> returns partial row, the subsequent scanner#next is supposed to return
> remaining cells. When this happens, cell rowkeys returned by servers i.e. any
> coprocessors is expected to be in the scan boundary range so that server can
> reliably perform its validation and return remaining cells as expected.
> Phoenix client initiates serial or parallel scans from the aggregators based
> on the region boundaries and the scan boundaries are sometimes adjusted based
> on where optimizer provided key ranges, to include tenant boundaries, salt
> boundaries etc. After the client opens the scanner and performs scan
> operation, some of the coprocs return invalid rowkey for the following cases:
> # Grouped aggregate queries
> # Some Ungrouped aggregate queries
> # Offset queries
> # Dummy cells returned with empty rowkey
> # Update statistics queries
> # Uncovered Index queries
> # Ordered results at server side
> # ORDER BY DESC on rowkey
> # Global Index read-repair
> # Paging region scanner with HBase scanner reopen
> # ORDER BY on non-pk column(s) with/without paging
> # GROUP BY on non-pk column(s) with/without paging
> Since many of these cases return reserved rowkeys, they are likely not going
> to match scan or region boundaries. It has potential to cause data integrity
> issues in certain scenarios as explained above. Empty rowkey returned by
> server can be treated as end of the region scan by HBase client.
> With the paging feature enabled, if the page size is kept low, we have higher
> chances of scanners returning dummy cell, resulting in increased num of RPC
> calls for better latency and timeouts. We should return only valid rowkey in
> the scan range for all the cases where we perform above mentioned operations
> like complex aggregate or offset queries etc.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
