Given your data model, there’s two ways you may read a tombstone:
You select an expired row, or you scan the whole table.
If you select an expired row, you’re going to scan one tombstone. With
sufficiently high read rate, that’ll look like you’re scanning a lot - each
read will add one to the histogram and it may add up to millions in 5 minutes
if you’re reading fast enough, but in this read pattern it’s not a problem.
If you’re doing a table scan, and you ask for 5000 rows at a time, you may have
to scan past tens of thousands of expired rows to eventually find the 5000 live
rows. IF you’re doing this, it may be a bit concerning, because it’s having to
skip past a ton of tombstones on the read path - which is expensive; this is
why the metric exists, but you’ve said you’re not doing this.
You’re not going to be able to stop reading tombstones unless you can stop the
app from reading expired rows. But on the plus side, this type of tombstone
read is not expensive and not concerning at all.
--
Jeff Jirsa
> On Feb 24, 2019, at 5:36 AM, Rahul Reddy <rahulreddy1...@gmail.com> wrote:
>
> Thanks Jeff. I'm trying to figure out why the tombstones scans are happening
> if possible eliminate it.
>
>> On Sat, Feb 23, 2019, 10:50 PM Jeff Jirsa <jji...@gmail.com> wrote:
>> G1GC with an 8g heap may be slower than CMS. Also you don’t typically set
>> new gen size on G1.
>>
>> Again though - what problem are you solving here? If you’re serving reads
>> and sitting under 50% cpu, it’s not clear to me what you’re trying to fix.
>> Tombstones scanned won’t matter for your table, so if that’s your only
>> concern, I’d ignore it.
>>
>>
>>
>> --
>> Jeff Jirsa
>>
>>
>>> On Feb 23, 2019, at 7:26 PM, Rahul Reddy <rahulreddy1...@gmail.com> wrote:
>>>
>>> ```jvm setting
>>>
>>> -XX:+UseThreadPriorities
>>> -XX:ThreadPriorityPolicy=42
>>> -XX:+HeapDumpOnOutOfMemoryError
>>> -Xss256k
>>> -XX:StringTableSize=1000003
>>> -XX:+AlwaysPreTouch
>>> -XX:-UseBiasedLocking
>>> -XX:+UseTLAB
>>> -XX:+ResizeTLAB
>>> -XX:+UseNUMA
>>> -XX:+PerfDisableSharedMem
>>> -Djava.net.preferIPv4Stack=true
>>> -XX:+UseG1GC
>>> -XX:G1RSetUpdatingPauseTimePercent=5
>>> -XX:MaxGCPauseMillis=500
>>> -XX:+PrintGCDetails
>>> -XX:+PrintGCDateStamps
>>> -XX:+PrintHeapAtGC
>>> -XX:+PrintTenuringDistribution
>>> -XX:+PrintGCApplicationStoppedTime
>>> -XX:+PrintPromotionFailure
>>> -XX:+UseGCLogFileRotation
>>> -XX:NumberOfGCLogFiles=10
>>> -XX:GCLogFileSize=10M
>>>
>>> Total memory
>>> free
>>> total used free shared buffers cached
>>> Mem: 16434004 16125340 308664 60 172872 5565184
>>> -/+ buffers/cache: 10387284 6046720
>>> Swap: 0 0 0
>>>
>>> Heap settings in cassandra-env.sh
>>> MAX_HEAP_SIZE="8192M"
>>> HEAP_NEWSIZE="800M"
>>> ```
>>>
>>>> On Sat, Feb 23, 2019, 10:15 PM Rahul Reddy <rahulreddy1...@gmail.com>
>>>> wrote:
>>>> Thanks Jeff,
>>>>
>>>> Since low writes and high reads most of the time data in memtables only.
>>>> When I noticed intially issue no stables on disk everything in memtable
>>>> only.
>>>>
>>>>> On Sat, Feb 23, 2019, 10:01 PM Jeff Jirsa <jji...@gmail.com> wrote:
>>>>> Also given your short ttl and low write rate, you may want to think about
>>>>> how you can keep more in memory - this may mean larger memtable and high
>>>>> flush thresholds (reading from the memtable), or perhaps the partition
>>>>> cache (if you are likely to read the same key multiple times). You’ll
>>>>> also probably win some with basic perf and GC tuning, but can’t really do
>>>>> that via email. Cassandra-8150 has some pointers.
>>>>>
>>>>> --
>>>>> Jeff Jirsa
>>>>>
>>>>>
>>>>>> On Feb 23, 2019, at 6:52 PM, Jeff Jirsa <jji...@gmail.com> wrote:
>>>>>>
>>>>>> You’ll only ever have one tombstone per read, so your load is based on
>>>>>> normal read rate not tombstones. The metric isn’t wrong, but it’s not
>>>>>> indicative of a problem here given your data model.
>>>>>>
>>>>>> You’re using STCS do you may be reading from more than one sstable if
>>>>>> you update column2 for a given column1, otherwise you’re probably just
>>>>>> seeing normal read load. Consider dropping your compression chunk size a
>>>>>> bit (given the sizes in your cfstats I’d probably go to 4K instead of
>>>>>> 64k), and maybe consider LCS or TWCS instead of STCS (Which is
>>>>>> appropriate depends on a lot of factors, but STCS is probably causing a
>>>>>> fair bit of unnecessary compactions and probably is very slow to expire
>>>>>> data).
>>>>>>
>>>>>> --
>>>>>> Jeff Jirsa
>>>>>>
>>>>>>
>>>>>>> On Feb 23, 2019, at 6:31 PM, Rahul Reddy <rahulreddy1...@gmail.com>
>>>>>>> wrote:
>>>>>>>
>>>>>>> Do you see anything wrong with this metric.
>>>>>>>
>>>>>>> metric to scan tombstones
>>>>>>> increase(cassandra_Table_TombstoneScannedHistogram{keyspace="mykeyspace",Table="tablename",function="Count"}[5m])
>>>>>>>
>>>>>>> And sametime CPU Spike to 50% whenever I see high tombstone alert.
>>>>>>>
>>>>>>>> On Sat, Feb 23, 2019, 9:25 PM Jeff Jirsa <jji...@gmail.com> wrote:
>>>>>>>> Your schema is such that you’ll never read more than one tombstone per
>>>>>>>> select (unless you’re also doing range reads / table scans that you
>>>>>>>> didn’t mention) - I’m not quite sure what you’re alerting on, but
>>>>>>>> you’re not going to have tombstone problems with that table / that
>>>>>>>> select.
>>>>>>>>
>>>>>>>> --
>>>>>>>> Jeff Jirsa
>>>>>>>>
>>>>>>>>
>>>>>>>>> On Feb 23, 2019, at 5:55 PM, Rahul Reddy <rahulreddy1...@gmail.com>
>>>>>>>>> wrote:
>>>>>>>>>
>>>>>>>>> Changing gcgs didn't help
>>>>>>>>>
>>>>>>>>> CREATE KEYSPACE ksname WITH replication = {'class':
>>>>>>>>> 'NetworkTopologyStrategy', 'dc1': '3', 'dc2': '3'} AND
>>>>>>>>> durable_writes = true;
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> ```CREATE TABLE keyspace."table" (
>>>>>>>>> "column1" text PRIMARY KEY,
>>>>>>>>> "column2" text
>>>>>>>>> ) WITH bloom_filter_fp_chance = 0.01
>>>>>>>>> AND caching = {'keys': 'ALL', 'rows_per_partition': 'NONE'}
>>>>>>>>> AND comment = ''
>>>>>>>>> AND compaction = {'class':
>>>>>>>>> 'org.apache.cassandra.db.compaction.SizeTieredCompactionStrategy',
>>>>>>>>> 'max_threshold': '32', 'min_threshold': '4'}
>>>>>>>>> AND compression = {'chunk_length_in_kb': '64', 'class':
>>>>>>>>> 'org.apache.cassandra.io.compress.LZ4Compressor'}
>>>>>>>>> AND crc_check_chance = 1.0
>>>>>>>>> AND dclocal_read_repair_chance = 0.1
>>>>>>>>> AND default_time_to_live = 18000
>>>>>>>>> AND gc_grace_seconds = 60
>>>>>>>>> AND max_index_interval = 2048
>>>>>>>>> AND memtable_flush_period_in_ms = 0
>>>>>>>>> AND min_index_interval = 128
>>>>>>>>> AND read_repair_chance = 0.0
>>>>>>>>> AND speculative_retry = '99PERCENTILE';
>>>>>>>>>
>>>>>>>>> flushed table and took tsstabledump
>>>>>>>>> grep -i '"expired" : true' SSTables.txt|wc -l
>>>>>>>>> 16439
>>>>>>>>> grep -i '"expired" : false' SSTables.txt |wc -l
>>>>>>>>> 2657
>>>>>>>>>
>>>>>>>>> ttl is 4 hours.
>>>>>>>>>
>>>>>>>>> INSERT INTO keyspace."TABLE_NAME" ("column1", "column2") VALUES (?,
>>>>>>>>> ?) USING TTL(4hours) ?';
>>>>>>>>> SELECT * FROM keyspace."TABLE_NAME" WHERE "column1" = ?';
>>>>>>>>>
>>>>>>>>> metric to scan tombstones
>>>>>>>>> increase(cassandra_Table_TombstoneScannedHistogram{keyspace="mykeyspace",Table="tablename",function="Count"}[5m])
>>>>>>>>>
>>>>>>>>> during peak hours. we only have couple of hundred inserts and 5-8k
>>>>>>>>> reads/s per node.
>>>>>>>>> ```
>>>>>>>>>
>>>>>>>>> ```tablestats
>>>>>>>>> Read Count: 605231874
>>>>>>>>> Read Latency: 0.021268529760215503 ms.
>>>>>>>>> Write Count: 2763352
>>>>>>>>> Write Latency: 0.027924007871599422 ms.
>>>>>>>>> Pending Flushes: 0
>>>>>>>>> Table: name
>>>>>>>>> SSTable count: 1
>>>>>>>>> Space used (live): 1413203
>>>>>>>>> Space used (total): 1413203
>>>>>>>>> Space used by snapshots (total): 0
>>>>>>>>> Off heap memory used (total): 28813
>>>>>>>>> SSTable Compression Ratio: 0.5015090954531143
>>>>>>>>> Number of partitions (estimate): 19568
>>>>>>>>> Memtable cell count: 573
>>>>>>>>> Memtable data size: 22971
>>>>>>>>> Memtable off heap memory used: 0
>>>>>>>>> Memtable switch count: 6
>>>>>>>>> Local read count: 529868919
>>>>>>>>> Local read latency: 0.020 ms
>>>>>>>>> Local write count: 2707371
>>>>>>>>> Local write latency: 0.024 ms
>>>>>>>>> Pending flushes: 0
>>>>>>>>> Percent repaired: 0.0
>>>>>>>>> Bloom filter false positives: 1
>>>>>>>>> Bloom filter false ratio: 0.00000
>>>>>>>>> Bloom filter space used: 23888
>>>>>>>>> Bloom filter off heap memory used: 23880
>>>>>>>>> Index summary off heap memory used: 4717
>>>>>>>>> Compression metadata off heap memory used: 216
>>>>>>>>> Compacted partition minimum bytes: 73
>>>>>>>>> Compacted partition maximum bytes: 124
>>>>>>>>> Compacted partition mean bytes: 99
>>>>>>>>> Average live cells per slice (last five minutes): 1.0
>>>>>>>>> Maximum live cells per slice (last five minutes): 1
>>>>>>>>> Average tombstones per slice (last five minutes): 1.0
>>>>>>>>> Maximum tombstones per slice (last five minutes): 1
>>>>>>>>> Dropped Mutations: 0
>>>>>>>>>
>>>>>>>>> histograms
>>>>>>>>> Percentile SSTables Write Latency Read Latency Partition
>>>>>>>>> Size Cell Count
>>>>>>>>> (micros) (micros)
>>>>>>>>> (bytes)
>>>>>>>>> 50% 0.00 20.50 17.08
>>>>>>>>> 86 1
>>>>>>>>> 75% 0.00 24.60 20.50
>>>>>>>>> 124 1
>>>>>>>>> 95% 0.00 35.43 29.52
>>>>>>>>> 124 1
>>>>>>>>> 98% 0.00 35.43 42.51
>>>>>>>>> 124 1
>>>>>>>>> 99% 0.00 42.51 51.01
>>>>>>>>> 124 1
>>>>>>>>> Min 0.00 8.24 5.72
>>>>>>>>> 73 0
>>>>>>>>> Max 1.00 42.51 152.32
>>>>>>>>> 124 1
>>>>>>>>> ```
>>>>>>>>>
>>>>>>>>> 3 node in dc1 and 3 node in dc2 cluster. With instanc type aws ec2
>>>>>>>>> m4.xlarge
>>>>>>>>>
>>>>>>>>>> On Sat, Feb 23, 2019, 7:47 PM Jeff Jirsa <jji...@gmail.com> wrote:
>>>>>>>>>> Would also be good to see your schema (anonymized if needed) and the
>>>>>>>>>> select queries you’re running
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> --
>>>>>>>>>> Jeff Jirsa
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>> On Feb 23, 2019, at 4:37 PM, Rahul Reddy <rahulreddy1...@gmail.com>
>>>>>>>>>>> wrote:
>>>>>>>>>>>
>>>>>>>>>>> Thanks Jeff,
>>>>>>>>>>>
>>>>>>>>>>> I'm having gcgs set to 10 mins and changed the table ttl also to 5
>>>>>>>>>>> hours compared to insert ttl to 4 hours . Tracing on doesn't show
>>>>>>>>>>> any tombstone scans for the reads. And also log doesn't show
>>>>>>>>>>> tombstone scan alerts. Has the reads are happening 5-8k reads per
>>>>>>>>>>> node during the peak hours it shows 1M tombstone scans count per
>>>>>>>>>>> read.
>>>>>>>>>>>
>>>>>>>>>>>> On Fri, Feb 22, 2019, 11:46 AM Jeff Jirsa <jji...@gmail.com> wrote:
>>>>>>>>>>>> If all of your data is TTL’d and you never explicitly delete a
>>>>>>>>>>>> cell without using s TTL, you can probably drop your GCGS to 1
>>>>>>>>>>>> hour (or less).
>>>>>>>>>>>>
>>>>>>>>>>>> Which compaction strategy are you using? You need a way to clear
>>>>>>>>>>>> out those tombstones. There exist tombstone compaction sub
>>>>>>>>>>>> properties that can help encourage compaction to grab sstables
>>>>>>>>>>>> just because they’re full of tombstones which will probably help
>>>>>>>>>>>> you.
>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>> --
>>>>>>>>>>>> Jeff Jirsa
>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>>> On Feb 22, 2019, at 8:37 AM, Kenneth Brotman
>>>>>>>>>>>>> <kenbrot...@yahoo.com.invalid> wrote:
>>>>>>>>>>>>>
>>>>>>>>>>>>> Can we see the histogram? Why wouldn’t you at times have that
>>>>>>>>>>>>> many tombstones? Makes sense.
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>> Kenneth Brotman
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>> From: Rahul Reddy [mailto:rahulreddy1...@gmail.com]
>>>>>>>>>>>>> Sent: Thursday, February 21, 2019 7:06 AM
>>>>>>>>>>>>> To: user@cassandra.apache.org
>>>>>>>>>>>>> Subject: Tombstones in memtable
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>> We have small table records are about 5k .
>>>>>>>>>>>>>
>>>>>>>>>>>>> All the inserts comes as 4hr ttl and we have table level ttl 1
>>>>>>>>>>>>> day and gc grace seconds has 3 hours. We do 5k reads a second
>>>>>>>>>>>>> during peak load During the peak load seeing Alerts for tomstone
>>>>>>>>>>>>> scanned histogram reaching million.
>>>>>>>>>>>>>
>>>>>>>>>>>>> Cassandra version 3.11.1. Please let me know how can this
>>>>>>>>>>>>> tombstone scan can be avoided in memtable
