Re: ways to improve compaction
On May 14, 2010, at 11:09 AM, Wout Mertens wrote: Old thread I know, but I was wondering about a way to make compaction more fluid: On Dec 21, 2009, at 23:20 , Damien Katz wrote: I saw recently some issues people where having with compaction, and I thought I'd get some thoughts down about ways to improve the compaction code/experience. 1. Multi-process pipeline processing. Similar to the enhancements to the view indexing, there is opportunities for pipelining operations instead of the current read/write batch operations it does. This can reduce memory usage and make compaction faster. 2. Multiple disks/mount points. CouchDB could easily have 2 or more database dirs, and each time it compacts, it copies the new database file to another dir/disk/mountpoint. For servers with multiple disks this will greatly smooth the copying as the disk heads won't need to seek between reads and writes. 3. Better compaction algorithms. There are all sorts of clever things that could be done to make the compaction faster. Right now it rebuilds the database in a similar manner as if it would if it clients were bulk updating it. This was the simplest way to do it, but certainly not the fastest. There are a lot of ways to make this much more efficient, they just take more work. 4. Tracking wasted space. This can be used to determine threshold for compaction. We don't need to track with 100% accuracy how much disk space is being wasted, but it would be a big improvement to at least know how much disk space the raw docs take, and maybe calculate an estimate of the indexes necessary to support them in a freshly compacted database. 5. Better Low level file driver support. Because we are using the Erlang built-in file system drivers, we don't have access to a lot of flags. If we had our own drivers, one option we'd like to use is to not OS cache the reads and write during the compaction, it's unnecessary for compaction and it could completely consume the cache with rarely accessed data, evicting lots of recently used live data, greatly hurting performance of other databases. Anyway, just getting these thoughts out. More ideas and especially code welcome. How about 6. Store the databases in multiple files. Instead of one really big file, use several big chunk-files of fixed maximum length. One chunk-file is active and receives writes. Once that chunk-file grows past a certain size, for example 25MB, start a new file. Then, at compaction time, you can do the compaction one chunk-file at a time. Possible optimization: If a certain chunk-file has no outdated documents (or only a small %), leave it alone. I'm armchair-programming here, I have only a vague idea of what the on-disk format looks like, but this could allow continuous compaction, by only compacting (slowly) the completed chunk-files. Furthermore, it would allow spreading the database across multiple disks (since there are now multiple files per db), although one disk would still be receiving all the writes. A smart write scheduler could make sure different databases have different active disks. Possibly, multiple chunk-files could be active at the same time, providing all sorts of interesting failure scenarios ;-) Thoughts? Wout. Hi Wout, Robert Newson suggested the very same in the original thread. It's a solid idea, to be sure. In related work, there's COUCHDB-738 https://issues.apache.org/jira/browse/COUCHDB-738 I wrote a patch to change the internal database format that allows compaction to skip an extra lookup in the by_id tree. Its a huge win for write-once DBs with random docids -- something like a 6x improvement in compaction speed in one test. However, DBs with frequently edited documents become 35-40% larger pre- and post-compaction. Damien has proposed a better alternative in that thread which is a much bigger rewrite of the compaction algorithm. Best, Adam
[jira] Commented: (COUCHDB-753) Add config option for view compact dir
[ https://issues.apache.org/jira/browse/COUCHDB-753?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=12867547#action_12867547 ] Till Klampaeckel commented on COUCHDB-753: -- I admit, I haven't really thought this through. My issue is that sometimes people run out of disk space with compaction. You (not necessarily you or CouchDB) could do something like block writes etc. when a compact is about to replace the database dir. Expose something from the server via JSON? Add config option for view compact dir -- Key: COUCHDB-753 URL: https://issues.apache.org/jira/browse/COUCHDB-753 Project: CouchDB Issue Type: Improvement Components: Database Core Reporter: Till Klampaeckel CouchDB creates a foo.view.compact file in the view directory (view_index_dir) when you run compact against a view. I'd really like to be able to specify another directory where this .compact file is created and worked on. This is especially helpful when it's difficult to run compaction because you run out of disk space on the same device. -- This message is automatically generated by JIRA. - You can reply to this email to add a comment to the issue online.
[jira] Updated: (COUCHDB-762) Faster implementation of couch_file:pread_iolist
[ https://issues.apache.org/jira/browse/COUCHDB-762?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ] Adam Kocoloski updated COUCHDB-762: --- Attachment: 762-pread_iolist-v2.patch An even better patch which does exactly 2 pread() calls in all cases, even for MD5-prefixed terms. Here are updated timings, with this approach termed 'pread_iolist_3': 4 pread_iolist_bench:go(5000, 1, 1, pread_iolist). Median 96 90%103 95%109 99%153 ok 5 pread_iolist_bench:go(5000, 1, 1, pread_iolist2). Median 82 90%90 95%94 99%107 ok 6 pread_iolist_bench:go(5000, 1, 1, pread_iolist3). Median 71 90%78 95%81 99%93 ok Faster implementation of couch_file:pread_iolist Key: COUCHDB-762 URL: https://issues.apache.org/jira/browse/COUCHDB-762 Project: CouchDB Issue Type: Improvement Components: Database Core Affects Versions: 0.11 Environment: any Reporter: Adam Kocoloski Priority: Minor Fix For: 1.1 Attachments: 762-pread_iolist-v2.patch, 762-pread_iolist.patch, patch-to-reproduce-benchmarks.txt, pread_iolist_bench.erl, pread_iolist_results.txt couch_file's pread_iolist function is used every time we read anything from disk. It makes 2-3 gen_server calls to the couch_file process to do its work. This patch moves the work done by the read_raw_iolist function into the gen_server itself and adds a pread_iolist handler. This means that one gen_server call is sufficient in every case. Here are some benchmarks comparing the current method with the patch that reduces everything to one call. I write a number of 10k binaries to a file, then read them back in a random order from 1/5/10/20 concurrent reader processes. I report the median/90/95/99 percentile response times in microseconds. In almost every case the patch is an improvement. The data was fully cached for these tests; I think that in a real-world concurrent reader scenario the performance improvement may be greater. The patch ensures that the 2-3 pread calls reading sequential bits of data (term length, MD5, and term) are always submitted without interruption. Previously, two concurrent readers could race to read different terms and cause some extra disk head movement. -- This message is automatically generated by JIRA. - You can reply to this email to add a comment to the issue online.
[jira] Closed: (COUCHDB-762) Faster implementation of couch_file:pread_iolist
[ https://issues.apache.org/jira/browse/COUCHDB-762?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ] Adam Kocoloski closed COUCHDB-762. -- Resolution: Fixed applied v2 of the patch Faster implementation of couch_file:pread_iolist Key: COUCHDB-762 URL: https://issues.apache.org/jira/browse/COUCHDB-762 Project: CouchDB Issue Type: Improvement Components: Database Core Affects Versions: 0.11 Environment: any Reporter: Adam Kocoloski Priority: Minor Fix For: 1.1 Attachments: 762-pread_iolist-v2.patch, 762-pread_iolist.patch, patch-to-reproduce-benchmarks.txt, pread_iolist_bench.erl, pread_iolist_results.txt couch_file's pread_iolist function is used every time we read anything from disk. It makes 2-3 gen_server calls to the couch_file process to do its work. This patch moves the work done by the read_raw_iolist function into the gen_server itself and adds a pread_iolist handler. This means that one gen_server call is sufficient in every case. Here are some benchmarks comparing the current method with the patch that reduces everything to one call. I write a number of 10k binaries to a file, then read them back in a random order from 1/5/10/20 concurrent reader processes. I report the median/90/95/99 percentile response times in microseconds. In almost every case the patch is an improvement. The data was fully cached for these tests; I think that in a real-world concurrent reader scenario the performance improvement may be greater. The patch ensures that the 2-3 pread calls reading sequential bits of data (term length, MD5, and term) are always submitted without interruption. Previously, two concurrent readers could race to read different terms and cause some extra disk head movement. -- This message is automatically generated by JIRA. - You can reply to this email to add a comment to the issue online.
[jira] Commented: (COUCHDB-762) Faster implementation of couch_file:pread_iolist
[ https://issues.apache.org/jira/browse/COUCHDB-762?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=12867602#action_12867602 ] Adam Kocoloski commented on COUCHDB-762: Just a comment about the breakdown of time spent reading a term. We're looking at median response times of 70 µs to read a 10K binary. I think this is roughly distributed as crypto:md5 - 30 µs pread()*2 - 20 µs everything else - µs Faster implementation of couch_file:pread_iolist Key: COUCHDB-762 URL: https://issues.apache.org/jira/browse/COUCHDB-762 Project: CouchDB Issue Type: Improvement Components: Database Core Affects Versions: 0.11 Environment: any Reporter: Adam Kocoloski Priority: Minor Fix For: 1.1 Attachments: 762-pread_iolist-v2.patch, 762-pread_iolist.patch, patch-to-reproduce-benchmarks.txt, pread_iolist_bench.erl, pread_iolist_results.txt couch_file's pread_iolist function is used every time we read anything from disk. It makes 2-3 gen_server calls to the couch_file process to do its work. This patch moves the work done by the read_raw_iolist function into the gen_server itself and adds a pread_iolist handler. This means that one gen_server call is sufficient in every case. Here are some benchmarks comparing the current method with the patch that reduces everything to one call. I write a number of 10k binaries to a file, then read them back in a random order from 1/5/10/20 concurrent reader processes. I report the median/90/95/99 percentile response times in microseconds. In almost every case the patch is an improvement. The data was fully cached for these tests; I think that in a real-world concurrent reader scenario the performance improvement may be greater. The patch ensures that the 2-3 pread calls reading sequential bits of data (term length, MD5, and term) are always submitted without interruption. Previously, two concurrent readers could race to read different terms and cause some extra disk head movement. -- This message is automatically generated by JIRA. - You can reply to this email to add a comment to the issue online.
[jira] Updated: (COUCHDB-704) Replication can lose checkpoints
[ https://issues.apache.org/jira/browse/COUCHDB-704?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ] Randall Leeds updated COUCHDB-704: -- Attachment: (was: rep-history-update-per-checkpoint.patch) Replication can lose checkpoints Key: COUCHDB-704 URL: https://issues.apache.org/jira/browse/COUCHDB-704 Project: CouchDB Issue Type: Bug Components: Replication Affects Versions: 0.12 Reporter: Randall Leeds Priority: Minor Attachments: save-all-rep-checkpoints.patch Original Estimate: 0h Remaining Estimate: 0h When saving replication checkpoints in the _local/repid document the new entry is always pushed onto the _original_ history list property that existed at the start of the replication. When any number of things causes the checkpoint to be written to only one of the databases the head of the history list gets out of sync. Subsequent attempts to start this replication must start from the latest common replication log entry in the _original_ history, as though this replication never occurred. A better idea is to push every checkpoint onto the history instead of replacing the head on each save. -- This message is automatically generated by JIRA. - You can reply to this email to add a comment to the issue online.
[jira] Commented: (COUCHDB-704) Replication can lose checkpoints
[ https://issues.apache.org/jira/browse/COUCHDB-704?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=12867724#action_12867724 ] Randall Leeds commented on COUCHDB-704: --- Looking over this once more. If we only append to the replication log, would it be more accurate to clear the stats after each checkpoint? The last log entry in the reply to a client request for non-continuous replication won't show the total number of documents replicated, but only the number since the last checkpoint. I don't know the best way to address this. Replication can lose checkpoints Key: COUCHDB-704 URL: https://issues.apache.org/jira/browse/COUCHDB-704 Project: CouchDB Issue Type: Bug Components: Replication Affects Versions: 0.12 Reporter: Randall Leeds Priority: Minor Attachments: save-all-rep-checkpoints.patch, whitespace.patch Original Estimate: 0h Remaining Estimate: 0h When saving replication checkpoints in the _local/repid document the new entry is always pushed onto the _original_ history list property that existed at the start of the replication. When any number of things causes the checkpoint to be written to only one of the databases the head of the history list gets out of sync. Subsequent attempts to start this replication must start from the latest common replication log entry in the _original_ history, as though this replication never occurred. A better idea is to push every checkpoint onto the history instead of replacing the head on each save. -- This message is automatically generated by JIRA. - You can reply to this email to add a comment to the issue online.
[jira] Updated: (COUCHDB-761) Timeouts in couch_log are masked, crashes callers
[ https://issues.apache.org/jira/browse/COUCHDB-761?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ] Randall Leeds updated COUCHDB-761: -- Attachment: improved-sync-logging.patch Here's my first run at a patch. Timeouts in couch_log are masked, crashes callers - Key: COUCHDB-761 URL: https://issues.apache.org/jira/browse/COUCHDB-761 Project: CouchDB Issue Type: Bug Components: Database Core Affects Versions: 0.10.1, 0.10.2, 0.11 Reporter: Randall Leeds Priority: Blocker Fix For: 0.10.3, 0.11.1, 1.0 Attachments: improved-sync-logging.patch Several users have reported seeing crash reports stemming from a function_clause match on handle_info in various gen_servers. The offending message looks like {#Ref, integer}. After months of banter and sleuthing, I determined that the likely cause was a late reply to a gen_server:call that timed out, with the #Ref being the tag on the response. After it came up again today in IRC, kocolosk quickly discovered that the problem appears to be in couch_log.erl. The logging macros (?LOG_*) call couch_log/*_on which calls get_level_integer/0. When this call times out the timeout is eaten and a late reply arrives to the calling process later, triggering the crash. Suggestions on how to fix this welcome. Ideas so far are async logging or infinite timeout. -- This message is automatically generated by JIRA. - You can reply to this email to add a comment to the issue online.