## Problem
There is no direct backpressure for compaction and L0 read
amplification. This allows a large buildup of compaction debt and read
amplification.
Resolves#5415.
Requires #10402.
## Summary of changes
Delay layer flushes based on the number of level 0 delta layers:
* `l0_flush_delay_threshold`: delay flushes such that they take 2x as
long (default `2 * compaction_threshold`).
* `l0_flush_stall_threshold`: stall flushes until level 0 delta layers
drop below threshold (default `4 * compaction_threshold`).
If either threshold is reached, ephemeral layer rolls also synchronously
wait for layer flushes to propagate this backpressure up into WAL
ingestion. This will bound the number of frozen layers to 1 once
backpressure kicks in, since all other frozen layers must flush before
the rolled layer.
## Analysis
This will significantly change the compute backpressure characteristics.
Recall the three compute backpressure knobs:
* `max_replication_write_lag`: 500 MB (based on Pageserver
`last_received_lsn`).
* `max_replication_flush_lag`: 10 GB (based on Pageserver
`disk_consistent_lsn`).
* `max_replication_apply_lag`: disabled (based on Pageserver
`remote_consistent_lsn`).
Previously, the Pageserver would keep ingesting WAL and build up
ephemeral layers and L0 layers until the compute hit
`max_replication_flush_lag` at 10 GB and began backpressuring. Now, once
we delay/stall WAL ingestion, the compute will begin backpressuring
after `max_replication_write_lag`, i.e. 500 MB. This is probably a good
thing (we're not building up a ton of compaction debt), but we should
consider tuning these settings.
`max_replication_flush_lag` probably doesn't serve a purpose anymore,
and we should consider removing it.
Furthermore, the removal of the upload barrier in #10402 will mean that
we no longer backpressure flushes based on S3 uploads, since
`max_replication_apply_lag` is disabled. We should consider enabling
this as well.
### When and what do we compact?
Default compaction settings:
* `compaction_threshold`: 10 L0 delta layers.
* `compaction_period`: 20 seconds (between each compaction loop check).
* `checkpoint_distance`: 256 MB (size of L0 delta layers).
* `l0_flush_delay_threshold`: 20 L0 delta layers.
* `l0_flush_stall_threshold`: 40 L0 delta layers.
Compaction characteristics:
* Minimum compaction volume: 10 layers * 256 MB = 2.5 GB.
* Additional compaction volume (assuming 128 MB/s WAL): 128 MB/s * 20
seconds = 2.5 GB (10 L0 layers).
* Required compaction bandwidth: 5.0 GB / 20 seconds = 256 MB/s.
### When do we hit `max_replication_write_lag`?
Depending on how fast compaction and flushes happens, the compute will
backpressure somewhere between `l0_flush_delay_threshold` or
`l0_flush_stall_threshold` + `max_replication_write_lag`.
* Minimum compute backpressure lag: 20 layers * 256 MB + 500 MB = 5.6 GB
* Maximum compute backpressure lag: 40 layers * 256 MB + 500 MB = 10.0
GB
This seems like a reasonable range to me.
Add wrappers for a few commands that didn't have them before. Move the
logic to generate tenant and timeline IDs from NeonCli to the callers,
so that NeonCli is more purely just a type-safe wrapper around
'neon_local'.
After #8655, we needed to mark some tests to shut down immediately. To
aid these tests, try the new pattern of `flush_ep_to_pageserver`
followed by a non-compacting checkpoint. This moves the general graceful
shutdown problem of having too much to flush at shutdown into the test.
Also, add logging for how long the graceful shutdown took, if we got to
complete it for faster log eyeballing.
Fixes: #8712
Cc: #8715, #8708
It seems that some benchmarks are failing because they are simply not
stopping to ingest wal on shutdown. It might mean that the tests were
never ran on a stable pageserver situation and WAL has always been left
to be ingested on safekeepers, but let's see if this silences the
failures and "stops the bleeding".
Cc: https://github.com/neondatabase/neon/issues/8712
## Problem
We reverted https://github.com/neondatabase/neon/pull/6661 a few days
ago. The change led to OOMs in
benchmarks followed by large WAL reingests.
The issue was that we removed [this
code](d04af08567/pageserver/src/tenant/timeline/walreceiver/walreceiver_connection.rs (L409-L417)).
That call may trigger a roll of the open layer due to
the keepalive messages received from the safekeeper. Removing it meant
that enforcing
of checkpoint timeout became even more lax and led to using up large
amounts of memory
for the in memory layer indices.
## Summary of changes
Piggyback on keep alive messages to enforce checkpoint timeout. This is
a hack, but it's exactly what
the current code is doing.
## Alternatives
Christhian, Joonas and myself sketched out a timer based approach
[here](https://github.com/neondatabase/neon/pull/6940). While discussing
it further, it became obvious that's also a bit of a hack and not the
desired end state. I chose not
to take that further since it's not what we ultimately want and it'll be
harder to rip out.
Right now it's unclear what the ideal system behaviour is:
* early flushing on memory pressure, or ...
* detaching tenants on memory pressure
This reverts commits 587cb705b8 (PR #6661)
and fcbe9fb184 (PR #6842).
Conflicts:
pageserver/src/tenant.rs
pageserver/src/tenant/timeline.rs
The conflicts were with
* pageserver: adjust checkpoint distance for sharded tenants (#6852)
* pageserver: add vectored get implementation (#6576)
Also we had to keep the `allowed_errors` to make `test_forward_compatibility` happy,
see the PR thread on GitHub for details.
587cb705b8
changed the layer rolling logic to more closely obey the
`checkpoint_distance` config. Previously, this test was getting
layers significantly larger than the 8K it was asking for. Now the
payload in the layers is closer to 8K (which means more layers in
total).
Tweak the `checkpoint_distance` to get a number of layers more
reasonable for this test. Note that we still get more layers than
before (~8K vs ~5K).
We use the term "endpoint" in for compute Postgres nodes in the web UI
and user-facing documentation now. Adjust the nomenclature in the code.
This changes the name of the "neon_local pg" command to "neon_local
endpoint". Also adjust names of classes, variables etc. in the python
tests accordingly.
This also changes the directory structure so that endpoints are now
stored in:
.neon/endpoints/<endpoint id>
instead of:
.neon/pgdatadirs/tenants/<tenant_id>/<endpoint (node) name>
The tenant ID is no longer part of the path. That means that you
cannot have two endpoints with the same name/ID in two different
tenants anymore. That's consistent with how we treat endpoints in the
real control plane and proxy: the endpoint ID must be globally unique.
Many python tests were setting the GC/compaction period to large
values, to effectively disable GC / compaction. Reserve value 0 to
mean "explicitly disabled". We also set them to 0 in unit tests now,
although currently, unit tests don't launch the background jobs at
all, so it won't have any effect.
Fixes https://github.com/neondatabase/neon/issues/2917
Replace the layer array and linear search with R-tree
So far, the in-memory layer map that holds information about layer
files that exist, has used a simple Vec, in no particular order, to
hold information about all the layers. That obviously doesn't scale
very well; with thousands of layer files the linear search was
consuming a lot of CPU. Replace it with a two-dimensional R-tree, with
Key and LSN ranges as the dimensions.
For the R-tree, use the 'rstar' crate. To be able to use that, we
convert the Keys and LSNs into 256-bit integers. 64 bits would be
enough to represent LSNs, and 128 bits would be enough to represent
Keys. However, we use 256 bits, because rstar internally performs
multiplication to calculate the area of rectangles, and the result of
multiplying two 128 bit integers doesn't necessarily fit in 128 bits,
causing integer overflow and, if overflow-checks are enabled,
panic. To avoid that, we use 256 bit integers.
Add a performance test that creates a lot of layer files, to
demonstrate the benefit.
