exp/lighthorizon/build/index-batch: merge map/reduce updates to latest on feature branch #4543
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Shaptic
approved these changes
Aug 16, 2022
| name: 'batch-map-job' | ||
| spec: | ||
| completions: 52 | ||
| parallelism: 10 |
There was a problem hiding this comment.
k8s job will only run 10 pods concurrently, and job will continue to start a new pod when one exits, until 52 pods have been launched, each pod getting assigned a unique JOB_COMPLETION_INDEX=0..51
| - name: FIRST_CHECKPOINT | ||
| value: "41426080" | ||
| - name: WORKER_COUNT | ||
| value: "8" |
There was a problem hiding this comment.
yes, a little bit of time slicing, these can be suggestive, for where this is aiming to deploy on dev cluster, there aren't huge amount of resources available, this also ties into the parallelism: 10 which is multiplier of cpu/ram limits here for job total.
There was a problem hiding this comment.
8 workers but 4 CPUs? 🤔
I think the process is mostly IO-bound so I think it's fine to have more workers than CPUs
| - name: BATCH_SIZE | ||
| value: "10048" | ||
| - name: FIRST_CHECKPOINT | ||
| value: "41426080" |
There was a problem hiding this comment.
yes, @2opremio mentioned running for roughly the last month to start with to verify the job stream, this is around July 1st.
| @@ -68,7 +68,7 @@ for (( i=0; i < $BATCH_COUNT; i++ )) | |||
| do | |||
| echo -n "Creating map job $i... " | |||
|
|
|||
| NETWORK_PASSPHRASE='testnet' MODULES='accounts_unbacked,transactions' \ | |||
| NETWORK_PASSPHRASE='testnet' JOB_INDEX_ENV='AWS_BATCH_JOB_ARRAY_INDEX' MODULES='accounts_unbacked,transactions' \ | |||
There was a problem hiding this comment.
Same comment here as on the other PR - why do we need the name of the name of the value rather than just passing the value itself? e.g. JOB_INDEX=$AWS_BATCH_JOB_ARRAY_INDEX?
There was a problem hiding this comment.
hmm, good point, this would be a good simplification, might squeeze it in.
There was a problem hiding this comment.
oh, I think I see why, under different orchestration environments like k8s, aws, that layer injects the job index number as a container env variable, the name can be anything, AWS_BATCH_JOB_ARRAY_INDEX or JOB_COMPLETION_INDEX for k8s, and there is no shell layer like here for interpolation, the only way the app code which got launched in the container deployed on orchestration layer can realistically reference/find the value is to be given the known deployment env var key name for job index instead and app code then performs the 2nd look up against that in env space to get the value.
| log.Errorf("Unable to mkdir %s, %v", dir, err) | ||
| return nil, err | ||
| } | ||
|
|
There was a problem hiding this comment.
This is a behavior change. I'm for it, but can we use this as a chance to add a docstring? e.g.
NewFileBackend connects to indices stored at `dir`, creating the directory if one doesn't
exist, and uses `parallel` to control how many workers to use when flushing to disk.
2opremio
reviewed
Aug 28, 2022
| -e START=5\ | ||
| -e END=50\ | ||
| -e END=150\ | ||
| -e NETWORK_PASSPHRASE="Test SDF Network ; September 2015"\ | ||
| -e CAPTIVE_CORE_CONFIG="/captive-core-testnet.cfg"\ | ||
| -e HISTORY_ARCHIVE_URLS="https://history.stellar.org/prd/core-testnet/core_testnet_001,https://history.stellar.org/prd/core-testnet/core_testnet_002"\ |
There was a problem hiding this comment.
I'm not sure it's a good idea to use testnet (since it can be reset)
There was a problem hiding this comment.
Also, there's a missing space before the horizon-light: job
There was a problem hiding this comment.
yes, intent was to avoid the long catchup seen on pubnet ranges, trade-off was a potential window of a couple minutes of time when newly reset testnet hasn't gen'd at least 150 ledgers yet during which this CI job step would fail if was invoked at same time.
I changed it on a small follow-up PR #4552 to reference same early range but on pubnet, maybe that suffices, need to check how long catchup runs on that, if it's quick, then maybe suffices?
2opremio
reviewed
Aug 28, 2022
| @@ -0,0 +1,7 @@ | |||
| # `stellar/horizon-indexer` | |||
There was a problem hiding this comment.
Isn't the image called stellar/lighthorizon-index-batch?
There was a problem hiding this comment.
yes, thanks for catch, fixed on small follow-up pr #4552.
Shaptic
added a commit
that referenced
this pull request
Jan 9, 2024
* exp/lighthorizon: Add initial support for XDR serialization (#4369) * exp/lighthorizon: Improve trie tests to avoid raw comparisons/outputs. (#4373) * exp/lighthorizon: Add XDR marshalling for the `TrieNode` structure. (#4375) * Add encoding stdlib interfaces * lighthorizon: Sync with upstream master branch (#4404) * services/ticker: ingest assets optimizations (#4218) * Add CHANGELOG entry for Horizon 2.14.0 release (#4208) (#4220) * Make sure we test reingestion for all possible operations (#4231) * services/horizon: Allow captive core to run with sqlite database (#4092) * services/horizon: Release DB connection in /paths when no longer needed (#4228) * services/horizon: Exclude trades with >10% rounding slippage from trade aggregations (#4178) * all: staticcheck fixes (#4239) * Migrate Horizon integration tests to GitHub Actions (#4242) * Fix StreamAllLiquidityPools and StreamAllOffers (#4236) * all: run builds and tests with go1.18rc1 (#4143) * all: cache go module downloads and other build and test artifacts (#3727) * services/horizon: Add LedgerHashStore to Captive-Core config (#4251) * all: migrate the rest of the CircleCI jobs to GitHub Actions (#4250) * horizon: Fix GitHub action problem with verify-range push in master (#4253) * all: fix ci ref_protected check for caching (#4254) * Switch over from CircleCI to GitHub A tions (#4256) * all: [GitHub actions] Reset the module and build cache in master/protected (#4266) * Forgot to add sudo in #4266 (#4270) * all: More go-setup github action fixes (#4274) * xdr: add instructions for generating xdr (#4280) * services/ticker: cache tomls during scraping (#4286) * services/ticker: use log fields during asset ingestion (#4288) * services/ticker: reduce size of toml cache in memory (#4289) * historyarchive: add --skip-optional flag (#3906) * all: Add Protocol 19 XDR and update StrKey to support Signed Payloads (#4279) * Replace keybase with publicnode in the stellar core config (#4291) * Fix captive core tests to write to /tmp, instead of polluting the repo (#4296) * all: remove go1.16 add go1.18 (#4284) * Rename methods and functions in submission system (#4298) * PR feedback (#4300) * Support new account fields for protocol-19. (#4294) * xdr, keypair: Add helpers to create CAP-40 decorated signatures (#4302) * services/horizon: Update txsub queue to account for new CAP-21 preconditions (#4301) * Uncomment StateVerifier test that generates account v3 extensions now that they are implemented. (#4304) * txnbuild: Add support for new CAP-21 preconditions. (#4303) * services/horizon: Support new CAP-21 transaction conditions (#4297) * txnbuild: Complete rename, avoid using XDR types in `TransactionParams`. (#4307) * all: Update Protocol 19 XDR to the latest (#4308) * services/horizon: Add a rate limit for path finding requests. (#4310) * clients/horizonclient: fix multi-parameter url for claimable balance query (#4248) * all: Fix Horizon integration tests (#4292) * horizon: Fix integration tests (#4314) * horizon: Set up protocol 19 integration tests infrastructure (#4312) * all: Change outdated CircleCI build badge (#4324) * horizon: Test new protocol 19 account fields (#4322) * all: update staticcheck to 2022.1 (#4326) * all: remove go.list and related docs (#4328) * horizon: Add transaction submission test for Protocol 19 (#4327) * Horizon v2.16.1 CHANGELOG (#4333) * Revert "Pin go versions temporarily" (#4338) * services/horizon: Use `bigint` over `timestamp` to accommodate large years (#4337) * xdr: Update xdrgen (#4341) * services/horizon: Change `min_account_sequence_age` column from `bigint` to string (#4339) * services/horizon: Bump stellar-core to v19.0.0rc1 for Horizon tests (#4345) * services/horizon: expose supported protocol version on root endpoint (#4347) * horizon: Small transaction submission refactoring (#4344) * services/horizon: Pass through nil ExtraSigners to avoid nil pointer deref (#4349) * doc: rename license file (#4350) * all: upgrade dep github.com/valyala/fasthttp (#4351) * services/horizon: Promote Stellar Core to v19.0.0 stable. (#4353) * services/horizon/integration: Precondition edge cases and V18->19 upgrade boundary. (#4354) * xdr: Synchronizes monorepo XDR with Stellar Core (#4355) * services/horizon: Properly allow nullable Protocol 19 account fields (#4357) * services/friendbot: include txhash in logs (#4359) * services/horizon: Improve transaction precondition `omitempty` behavior (#4360) * tools/horizon-cmp: Improve panic error message (#4365) * services/horizon: Merge stable v2.17.0 back into master: (#4363) * Use UNIX timestamps instead of RFC3339 strings for timebounds. (#4361) * xdrgen: remove gemfile and rakefile to just use docker for the xdrgen (#4366) * Conservatively limit the number of DB connections of integration tests (#4368) * internal/integrations: db_test should drop test db instances when finished (#4185) * GHA: Bump Core version to v19.0.1 in Horizon workflows. (#4378) * services/horizon, clients/horizonclient: Allow filtering ingested transactions by account or asset. (#4277) * Push stellar/ledger-state-diff images from Github actions (#4380) * services/horizon: Fixes copy-paste typo in `--help` text (#4383) * tools/alb-replay: Add new features to alb-replay (#4384) * services/horizon: Optimize claimable balances query to limit records earlier (#4385) * support/db, services/horizon/internal: Configure postgres client connection timeouts for read only db (#4390) * Refactor trade aggregation query. (#4389) * services/horizon/internal/db2/history: Implement StreamAllOffers using batches (#4397) * Add flag to disable path finding endpoints (#4399) Co-authored-by: stfung77 <[email protected]> Co-authored-by: Leigh McCulloch <[email protected]> Co-authored-by: Alfonso Acosta <[email protected]> Co-authored-by: Paul Bellamy <[email protected]> Co-authored-by: Bartek Nowotarski <[email protected]> Co-authored-by: tamirms <[email protected]> Co-authored-by: Alfonso Acosta <[email protected]> Co-authored-by: Graydon Hoare <[email protected]> Co-authored-by: Satyam Zode <[email protected]> Co-authored-by: Satyam Zode <[email protected]> Co-authored-by: erika-sdf <[email protected]> Co-authored-by: iateadonut <[email protected]> Co-authored-by: Shawn Reuland <[email protected]> Co-authored-by: shawn <[email protected]> Co-authored-by: Shivendra Mishra <[email protected]> Co-authored-by: Jacek Nykis <[email protected]> Co-authored-by: jacekn <[email protected]> * Explain map and reduce commands * exp/lighthorizon: Refactor single-process index builder. (#4410) * Refactor index builder: - allow worker count to be a command line parameter - split work by checkpoints rather than ledgers - move actual index insertion work to helpers - move progress bar into helpers - simplify participants code, payments vs. all * Properly work on a checkpoint range at a time: - previously, it was just arbitrary 64-ledger chunks which is not as helpful * Define a generic module processing function * Move index building into a separate object * Fix off-by-one error in checkpoint index builder: - Keeping this as-is would mean that the first chunk of ledgers will be "Checkpoint 0" which doesn't make sense in the bitmap - Calling index.setActive(0) is essentially a no-op, because no bit will ever be set. - In the case of an empty index in which the only active account checkpoint is the first one, this is indistinguishable from an index with no activity. * exp/services/ledgerexporter: Extend tool to support lower ledger bound. (#4405) * exp/lighthorizon: Refactor and repair the reduce job (#4424) * Use envvars for every configurable thing, incl. index sources and final merged index target: This removes any hard dependency on S3 and lets you use any supported backend for the map-reduce operation. It was done specifically with local filesystem-based testing in mind, but naturally opens up other backends as well. * Add lots of helper functions: Specifically, helpers now exist for both merging two sets of named indices together and partitioning work based on the account/transaction hashes into separate jobs/routines. * Lots more logging! For progress tracking, debugging, etc. * Create a thread-safe string set abstraction for tracking completed work. * Better error handling: `os.IsNotExist(err)` is much more reliable over a direct equality check to `ErrNotExist`. This also ties in to backend-independence. We can also log and return an error rather than immediately panicking on its occurrence. * Transaction flushes need to be thread-safe if they're going to be done from different goroutines during reduction. Otherwise, you get panics from concurrent writes to its maps. * The "account list" (aka the file containing a list of all accounts in the partitioned index) needs to be flushed at the same time as the index itself: If this isn't done, then `FlushAccounts()` will do absolutely nothing after a `Flush()`, because the previous `Flush()` will clear the map of indices out of memory. Since the account list comes from memory, it becomes a no-op. * Split work across multiple channels rather than just one If the work comes from a single channel, accounts can get skipped overall because they aren't put back on the queue if they're skipped by a single worker. It makes more sense to make each worker have its own channel, partitioning the work *before* it gets to the worker rather than after. * exp/lighthorizon: Unify map-reduce and single-process index builders (#4423) * Main thing: `./index/cmd/single` and `./index/cmd/batch/map` now leverage the same index building code (i.e. `BuildIndices`) * This also extends the map-reduce builder to take the txmeta source / index destination URLs from envvars rather: This eliminates a hard dependency on S3, and it's done here because splitting that out from the giga-PR was difficult. * We can infer checkpoints from `ledger.LedgerSequence()` rather than passing them in as a parameter, which cleans up modules. * This finally adds a new `ProcessAccountsWithoutBackend` module for the Map job * exp/lighthorizon: Thread-safe support for reading account list via FileBackend (#4422) Three key changes: - actually read the account list when using a filesystem backend - using `O_APPEND` on the file to support concurrent writes - ensure that the read list is a unique set of accounts * exp/lighthorizon: Restructure index package into sensible sub-packages (#4427) * exp/lighthorizon: Merge on-disk index with in-memory one on load. (#4435) * Add test for single-process index builder * Merge in-memory index with on-disk one when loading * Add fixture of unpacked ledgers for fast local testing * Isolate the index we need to merge * Use a ByteReader so that multiple indices in one file work 🤦 * Add to/from XDR support to bitmap index * Fix and extend gzip tests to handle the bytereader bug * Simplify participant processing code * exp/lighthorizon: Allow indexer to continually update as new txmeta appears (#4432) * exp/lighthorizon: enforce the limit from request on the response size (#4431) * Dockerize ledgerexport to run in AWS Batch This Change: 1. Creates docker image (stellar/horizon-ledgerexporter) which works in a similar fashion to stellar/horizon-verify-rage and is tested and pushed as part of the Horizon GitHub workflow. 2. Adds two more parameters to ledgerexporter * --end-ledger: which indicates at what ledger to stop the export * --write-latest-path: which indicates whether to udpate the /latest path of the target Latest path writing is disabled in the container by default in order to avoid race-conditions between parallel jobs * exp/lighthorizon: Add test for batch index building map job (#4440) * Modify single-process test to generalize to whatever fixture data exists This also adds a test to check that single-process works on a non-checkpoint starting point which is important. * Fix map program to properly build sub-paths depending on its job index Previously, this only happened for explicitly S3 backends. * Make map job default to using all CPUs * Stop clearing indices from memory if using unbacked module * Use historyarchive.CheckpointManager for all checkpoint math * Update lastBuiltLedger w/ safely concurrent writes * Refactor bound preparation and add --continue flag * Address review feedback and rework env variable names * Run gofmt -w (I don't know why those files were changed) * Add proper logging to indicate what range is being exported * Add clarification about end ledger * Fix boolean argument passing * Address review feedback * Address feedback * Use sqlite for captive core * exp/lighthorizon: Add basic scaffolding for metrics. (#4456) * Use correct network passphrase when populating transaction * Add scaffolding for Prom/log metrics and some example ones * Misc. clarifications and fixes to the index builder * lighthorizon: Prepend version to ledger files (#4450) * Prepend version to ledger files * Encode versioning in XDR * Regenerate fixtures * Fix ledger fixtures * Appease govet * Move all lighthorizon types to /xdr * exp/lighthorizon/index: More testing for batch indexing and off-by-one bugfix. (#4442) * Add reduce test to ensure combining map jobs works * Actually test that TOIDs are correct * Bugfix: Transaction prefix loop should be inclusive * Isolate loggers to individual processing "sections" * Minor ledgerexporter infrastructure improvements (#4461) * Push the stellar/horizon-ledgerexporter docker image when pushing to the lighthorizon branch * Fix the ledger exporter aws batch jobs when running on the first batch * Forgot to add login step to ledgerexporter workflow * exp/lighthorizon: Set a default number of workers. (#4465) * Default to the number of CPUs if worker count isn't specified * Set a timeout on the reduce job to avoid test suite hanging indefinitely * exp/lighthorizon: Fix the single-process index builder data race. (#4470) * Add synchronization for the work submission routine. Thank you @sreuland! Co-authored-by: shawn <[email protected]> * /exp/lighthorizon: new endpoints for tx and ops paged listing by account id (#4453) * exp/lighthorizon: Add an on-disk cache for frequently accessed ledgers. (#4457) * Replace custom LRU solution with an off-the-shelf data structure. * Add a filesystem cache in front of the ledger backend to lower latency * Add cache size parameter; only setup cache if not file:https:// * Extract S3 region from the archive URL if it's applicable. * exp/lighthorizon/index: Drop building indices for successful transactions. (#4482) * Add metrics middleware to collect request duration metrics (#4486) * exp/lighthorizon: Isolate cursor advancement code to its own interface (#4484) * Move cursor manipulation code to a separate interface * Small test refactor to improve readability and long-running lines * Combine tx and op tests into subtests * Fix how IndexStore is mocked out * exp/lighthorizon/index: Parse network passphrase from the env. (#4491) * Refactor access to meta archive (#4488) Refactor `historyarchive` and `ledgerbackend` to allow better access to the new meta archives: * Created `metaarchive` package that connects to the new meta archives (and allows accessing `xdr.SerializedLedgerCloseMeta`). * Extracted `ArchiveBackend` to the new `support/storage` package as it contains only storage related methods. New package is used in both `historyarchive` and `metaarchive`. * exp/lighthorizon: Add response age prometheus metrics (#4492) * exp/lighthorizon/index: Allow accounts to be indexed by ledger. (#4495) * Add builders to make account indices by ledger * Add `MODULE` parameter to map job in batch builder * Don't build transaction indices by default * services/horizon/docker/ledgerexporter: deploy ledgerexporter image as service (#4490) * Make indexing s3 bucket configurable (#4507) * exp/lighthorizon: Add duration metrics for on-the-fly ingestion elements. (#4476) Add basic aggregate metrics for request fulfillment: - how long did ledger downloads take, on average? - how long did ledger processing take, on average? - how long did index lookups take, on average? - how many ledgers were needed? - how long did the entire request take, in total? * exp/lighthorizon: Add JSON content type to responses. (#4509) * exp/lighthorizon: *Correctly* set `Content-Type`, plus JSONify errors (#4513) * exp/lighthorizon/services: Move service-specific stuff to its own file. (#4502) * exp/lighthorizon, xdr: Rename `CheckpointIndex` to better reflect its capabilty. (#4510) * Rename NextActive -> NextActiveBit to be descriptive * exp/lighthorizon: Add a suite of tools to manage the on-disk ledger cache. (#4522) * Run 'go mod tidy' after merge * exp/lighthorizon: add horizon web docker/k8s deployment (#4519) * It seems like the merge caused some deleted files to stay in: The commit b3407fd from PR #4418 deleted these files, so we just do the same. A quick manual inspection showed us that the deltas transferred over, just not the deletions, for some reason. Idk why these changes ended up in the code, kinda sus... More deleted files snuck in? * One more that didn't get removed 🤔 * all: Incorporate generics into Light Horizon code. (#4537) * bump go version to 18 on lighthorizon docker images, they need it now (#4541) * exp/lighthorizon/actions: use standard Problem model on API error responses (#4542) * exp/lighthorizon/build/index-batch: carry over map/reduce updates to latest docker layout on feature branch (#4543) * exp/lighthorizon: Properly transform transactions into JSON. (#4531) * exp/lighthorizon: Add a set of tools to aide in index inspection. (#4561) * exp/lighthorizon/cmd: index batch fix s3 sub paths in reduce (#4552) * exp/lighthorzon: Add a generic, thread-safe `SafeSet`. (#4572) * support/storage: Make the on-disk cache thread-safe. (#4575) * exp/lighthorizon: Incorporate tool subcommands into the webserver. (#4579) * exp/lighthorizon/index/cmd: Fix index single watch, slow down the retry on not-found ledgers (#4582) * exp/lighthorizon: Refactor archive interface and support parallel ledger downloads. (#4548) - Refactor and simplify Archive abstraction to incorporate MetaArchive - Actually add & use parallel downloads, preparing checkpoint chunks - Fix test structures and mocking - Fix cache to ignore on-disk if lockfile present * exp/lighthorizon: Minor error-handling and deployment improvements. (#4599) - actually set the PARALLEL_DOWNLOADS parameter to use #4468 - return a 404 rather than a 500 if a ledger is missing as its more descriptive - handle `count = 0` in average metric calculations * exp/lighthorizon/index: Add ability to disable bits in index. (#4601) * exp/lighthorizon: Add parameters to preload ledger cache. (#4615) * Add ability to preload cache in parallel after launching webserver * Default to 1 day of ledgers @ 6s each --------- Co-authored-by: Bartek Nowotarski <[email protected]> Co-authored-by: Paul Bellamy <[email protected]> Co-authored-by: Bartek <[email protected]> Co-authored-by: Bartek <[email protected]> Co-authored-by: tamirms <[email protected]> Co-authored-by: George <[email protected]> Co-authored-by: stfung77 <[email protected]> Co-authored-by: Leigh McCulloch <[email protected]> Co-authored-by: Alfonso Acosta <[email protected]> Co-authored-by: Alfonso Acosta <[email protected]> Co-authored-by: Graydon Hoare <[email protected]> Co-authored-by: Satyam Zode <[email protected]> Co-authored-by: Satyam Zode <[email protected]> Co-authored-by: erika-sdf <[email protected]> Co-authored-by: iateadonut <[email protected]> Co-authored-by: Shawn Reuland <[email protected]> Co-authored-by: shawn <[email protected]> Co-authored-by: Shivendra Mishra <[email protected]> Co-authored-by: Jacek Nykis <[email protected]> Co-authored-by: jacekn <[email protected]> Co-authored-by: George Kudrayvtsev <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
PR Checklist
PR Structure
otherwise).
services/friendbot, orallordocif the changes are broad or impact manypackages.
Thoroughness
.mdfiles, etc... affected by this change). Take a look in the
docsfolder for a given service,like this one.
Release planning
needed with deprecations, added features, breaking changes, and DB schema changes.
semver, or if it's mainly a patch change. The PR is targeted at the next
release branch if it's not a patch change.
What
This is merging the net changes to map/reduce from #4501 to the latest on main feature branch, lighthorizon, which had changes merged from #4519 which refactored/consolidated related docker images.
Why
remove merge conflicts so net changes to map/reduce can be be committed, this PR represents the net changes needed from #4501 and deprecates need to merge that PR.
Known limitations